Skip to main content

rudb_encoding/
string.rs

1//! The string column, which is offsets, bytes, and the choice between compressing the bytes and
2//! not storing most of them at all.
3//!
4//! ClickBench `hits` is a string dataset before it is anything else. `URL`, `Referer`, `Title` and
5//! the referer derived columns are most of the 20.46 GB DuckDB writes for it, so most of what
6//! `spec/02-the-goal.md` promises on the resource axis has to come out of this file.
7//!
8//! ## The five shapes
9//!
10//! `CONSTANT` when every value is the same. `PLAIN`, which is lengths and raw bytes and is the
11//! baseline the others have to beat. `FSST`, which is a symbol table and the same lengths over
12//! compressed bytes. `DICT`, which is the distinct values and an array of codes. `FRONT`, which is
13//! the length of the prefix each value shares with the one before it and the rest of the value.
14//!
15//! `DICT_FSST` from the section 6.2 table is not a sixth shape. A dictionary's entries are a string
16//! column, and encoding them goes back through the same chooser, so a dictionary whose entries are
17//! FSST compressed is what the chooser produces on its own whenever that is smaller. The same
18//! recursion gives run length encoding of strings for free, because the codes are an integer chunk
19//! and `crate::integer` already knows what to do with a column of long runs.
20//!
21//! ## Why front coding is here
22//!
23//! The whole file measurement in M1 says the chooser produces 11.65 GB for `hits` against Parquet's
24//! 13.76 GB, and that `URL`, `Referer` and `OriginalURL` are 6.11 GB of it, and that on those three
25//! the chooser loses to Parquet's Snappy. The shape it picked on all three was `DICT(FSST[255])`,
26//! so the cascade was working and FSST was still losing.
27//!
28//! The reason is structural. FSST compresses each value on its own against a 255 symbol table, and
29//! a block compressor has the previous few kilobytes of the page to point back into. Two URLs that
30//! share a host and half a path are most of a back reference to each other and are nothing at all
31//! to a symbol table, which can only spend eight bytes of a symbol on the part they share and has
32//! to spend it again on every value. On a sorted dictionary of URLs the value before is the closest
33//! thing in the column to the value in hand, and the bytes they share are the redundancy Snappy was
34//! finding. Front coding is what reaches those bytes, and it composes with everything else here:
35//! the suffixes it leaves behind are a string column and go back through the chooser, so
36//! `DICT(FRONT(FSST))` is a shape the chooser can arrive at without anyone naming it.
37//!
38//! The chain has no restarts, so reading entry `n` means walking from entry zero. That is the right
39//! trade while a dictionary is decoded whole, which is what `decode` does. When something wants one
40//! entry out of a dictionary without materialising the rest, the answer is a restart every so many
41//! entries, and it costs one full value per block.
42//!
43//! ## Lengths, not offsets
44//!
45//! The usual layout is `n + 1` offsets and Arrow does it that way because a slice of an array has
46//! to be free. On disk the offsets are a monotonically increasing sequence whose differences are
47//! the lengths, and the differences are what compress: URL lengths in a real column are a few dozen
48//! distinct values in a narrow band, which the integer cascade turns into a handful of bits each,
49//! while the offsets themselves need enough bits to address the whole chunk. The integer cascade
50//! would find that by choosing DELTA, and storing lengths directly gets to the same place without
51//! spending a level of the cascade on it. Offsets are a prefix sum away and that is a decode time
52//! cost of one add per value.
53//!
54//! ## What is not here
55//!
56//! Nulls. A chunk here is N byte strings and an empty string is a value like any other. Validity is
57//! a bitmap that belongs to the column rather than to the encoding, per `spec/05-storage.md`, and
58//! `ROARING` in the section 6.2 table is what encodes it.
59//!
60//! Shared symbol tables and shared dictionaries across columns, which are section 6.4 and are the
61//! measurement this milestone exists for. Everything here is one column on its own, which is the
62//! baseline they get compared against.
63
64use rudb_common::{Error, Result};
65
66use crate::chooser::{Chooser, EXHAUSTIVE, Settled};
67use crate::fsst::SymbolTable;
68use crate::integer;
69use crate::lz;
70use crate::reader::Reader;
71
72/// How deep the recursion goes. A dictionary of a dictionary is not a thing, so this only has to
73/// stop the dictionary's own entries from being dictionary encoded again.
74const MAX_DEPTH: u8 = 2;
75
76/// How little sharing between neighbours is still worth offering front coding for, as one over
77/// this. A twentieth of the column is around where the prefix lengths start paying for themselves,
78/// and below it the candidate is an encode of the whole column that loses.
79const SHARE_DIVISOR: usize = 20;
80
81/// How few bytes is too few to bother looking for repeats in.
82///
83/// The matcher costs a hash table and a pass over the bytes whether it wins or not, and the chooser
84/// is exhaustive, so an ungated candidate is a tax on every string column in the database. Four
85/// kilobytes is about where a 32 KiB window has enough behind it to find anything.
86const LZ_FLOOR: usize = 4096;
87
88/// How many bytes of a column the symbol table is trained on.
89///
90/// The paper trains on about 16 KB. This is four times that, because training happens once per
91/// chunk here rather than once per block, and because the cost of a symbol that is only in the
92/// sample by accident is paid on every value in the chunk.
93pub(crate) const SAMPLE_BYTES: usize = 64 * 1024;
94
95/// What a string chunk is encoded as. The discriminant is the tag byte and is part of the format.
96#[derive(Debug, Clone, Copy, PartialEq, Eq)]
97pub enum Kind {
98    /// One value repeated.
99    Constant = 0,
100    /// Lengths and raw bytes.
101    Plain = 1,
102    /// Lengths, a symbol table, and FSST compressed bytes.
103    Fsst = 2,
104    /// The distinct values as a string chunk of their own, and codes into it as an integer chunk.
105    Dict = 3,
106    /// Shared prefix lengths as an integer chunk, and what is left of each value as a string chunk.
107    Front = 4,
108    /// Value lengths, copy lengths and copy offsets as integer chunks, and the bytes no copy
109    /// covered as a string chunk. See the `lz` module for what the matcher does and why it is here.
110    Lz = 5,
111}
112
113impl Kind {
114    fn tag(self) -> u8 {
115        self as u8
116    }
117
118    fn from_tag(tag: u8) -> Result<Self> {
119        match tag {
120            0 => Ok(Self::Constant),
121            1 => Ok(Self::Plain),
122            2 => Ok(Self::Fsst),
123            3 => Ok(Self::Dict),
124            4 => Ok(Self::Front),
125            5 => Ok(Self::Lz),
126            other => Err(Error::internal(format!("unknown string encoding tag {other}"))),
127        }
128    }
129
130    /// The name that goes in a report.
131    #[must_use]
132    pub fn name(self) -> &'static str {
133        match self {
134            Self::Constant => "CONSTANT",
135            Self::Plain => "PLAIN",
136            Self::Fsst => "FSST",
137            Self::Dict => "DICT",
138            Self::Front => "FRONT",
139            Self::Lz => "LZ",
140        }
141    }
142}
143
144/// Encodes a chunk of strings, choosing whatever comes out smallest.
145///
146/// Every candidate that applies is encoded in full and the smallest is kept, which is what this has
147/// always done and is what every size this crate has reported came out of. [`encode_with`] is the
148/// same thing with the search made swappable.
149///
150/// # Errors
151///
152/// If the chunk is longer than `u32::MAX` values, or if an encoding produces something its own
153/// decoder would not accept.
154pub fn encode(values: &[&[u8]]) -> Result<Vec<u8>> {
155    encode_with(values, &EXHAUSTIVE)
156}
157
158/// [`encode`] with somebody else deciding which candidates are worth encoding in full.
159///
160/// A chooser narrows the list and nothing else. It cannot offer a candidate that does not apply, so
161/// whatever it picks still has to encode the whole chunk and still has to decode, and the worst a
162/// bad one can do is come out bigger than [`encode`] would have.
163///
164/// # Errors
165///
166/// As [`encode`].
167pub fn encode_with(values: &[&[u8]], chooser: &dyn Chooser) -> Result<Vec<u8>> {
168    encode_at(values, 0, chooser)
169}
170
171/// A decoded chunk as one buffer with the values laid end to end, and where each one ends in it.
172///
173/// This is what the decoder builds and [`decode`] is a copy out of it. The cascade is why: a nest
174/// like `FRONT(LZ(FSST))` decodes three levels to produce one, and a level that hands its caller a
175/// `Vec<Vec<u8>>` has allocated once per value and copied every byte it holds. Three levels of that
176/// on a chunk of a thousand URLs is three thousand allocations to produce a thousand strings that
177/// the caller almost always wants back to back anyway.
178///
179/// It also makes the levels cheaper on their own terms. `PLAIN` is one `memcpy` of the whole
180/// payload because the values are already end to end in the file. `FRONT` copies a shared prefix
181/// out of the buffer it is writing into, so the previous value never has to be somewhere else.
182/// `LZ` replays straight into the buffer, which is what its copy offsets meant in the first place.
183#[derive(Debug, Clone, Default, PartialEq, Eq)]
184pub struct Flat {
185    bytes: Vec<u8>,
186    /// Where each value ends, so a value starts where the one before it ended and the last entry
187    /// is the length of `bytes`. Ends rather than offsets because a value is appended and its end
188    /// is what is known at that moment.
189    ends: Vec<usize>,
190}
191
192impl Flat {
193    fn with_capacity(count: usize, bytes: usize) -> Self {
194        Self { bytes: Vec::with_capacity(bytes), ends: Vec::with_capacity(count) }
195    }
196
197    fn push(&mut self, value: &[u8]) {
198        self.bytes.extend_from_slice(value);
199        self.ends.push(self.bytes.len());
200    }
201
202    /// Where the value at `index` starts, which is where the one before it ended.
203    fn start(&self, index: usize) -> usize {
204        if index == 0 { 0 } else { self.ends[index - 1] }
205    }
206
207    /// How many values the chunk holds.
208    #[must_use]
209    pub fn len(&self) -> usize {
210        self.ends.len()
211    }
212
213    /// Whether the chunk holds no values at all, which is not the same as holding empty ones.
214    #[must_use]
215    pub fn is_empty(&self) -> bool {
216        self.ends.is_empty()
217    }
218
219    /// The values laid end to end. A caller that already knows the boundaries, which is what a
220    /// global dictionary's offsets are, needs nothing else.
221    #[must_use]
222    pub fn bytes(&self) -> &[u8] {
223        &self.bytes
224    }
225
226    /// The value at `index`, or `None` past the end.
227    #[must_use]
228    pub fn get(&self, index: usize) -> Option<&[u8]> {
229        let end = *self.ends.get(index)?;
230        self.bytes.get(self.start(index)..end)
231    }
232
233    /// Every value in order.
234    pub fn iter(&self) -> impl Iterator<Item = &[u8]> {
235        let mut at = 0;
236        self.ends.iter().map(move |end| {
237            let value = self.bytes.get(at..*end).unwrap_or_default();
238            at = *end;
239            value
240        })
241    }
242
243    /// The buffer on its own, for a caller that wanted the bytes rather than the values.
244    #[must_use]
245    pub fn into_bytes(self) -> Vec<u8> {
246        self.bytes
247    }
248
249    /// The buffer and the ends that divide it, for a caller building its own layout over them.
250    ///
251    /// [`into_bytes`](Self::into_bytes) is enough for a caller that already knows where the values
252    /// end, which is what a global dictionary's stored offsets are. A caller that does not know has
253    /// only [`iter`](Self::iter), and walking that to build a run of boundaries copies out numbers
254    /// the chunk already holds. This hands both halves over and keeps the one allocation each.
255    #[must_use]
256    pub fn into_parts(self) -> (Vec<u8>, Vec<usize>) {
257        (self.bytes, self.ends)
258    }
259
260    fn into_values(self) -> Vec<Vec<u8>> {
261        let mut values = Vec::with_capacity(self.len());
262        let mut at = 0;
263        for end in &self.ends {
264            values.push(self.bytes[at..*end].to_vec());
265            at = *end;
266        }
267        values
268    }
269}
270
271/// Decodes a chunk written by [`encode`] without taking it apart into a value each.
272///
273/// # Errors
274///
275/// As [`decode`].
276pub fn decode_flat(bytes: &[u8]) -> Result<Flat> {
277    let mut reader = Reader::new(bytes);
278    let flat = decode_chunk(&mut reader)?;
279    if reader.remaining() != 0 {
280        return Err(Error::internal(format!(
281            "{} bytes left over after decoding a string chunk",
282            reader.remaining()
283        )));
284    }
285    Ok(flat)
286}
287
288/// Decodes a chunk that sits at the front of a longer buffer, and says how many bytes it took.
289///
290/// A column group holds one of these per column, and the decoder on that side cannot know where
291/// one ends until it has been read.
292///
293/// # Errors
294///
295/// As [`decode`], except that trailing bytes are what the caller asked about rather than an error.
296pub fn decode_prefix(bytes: &[u8]) -> Result<(Vec<Vec<u8>>, usize)> {
297    let mut reader = Reader::new(bytes);
298    let values = decode_chunk(&mut reader)?;
299    Ok((values.into_values(), reader.used()))
300}
301
302/// [`describe`] over a chunk at the front of a longer buffer, and how many bytes it took.
303///
304/// # Errors
305///
306/// As [`decode_prefix`].
307pub fn describe_prefix(bytes: &[u8]) -> Result<(String, usize)> {
308    let mut reader = Reader::new(bytes);
309    let text = describe_chunk(&mut reader)?;
310    Ok((text, reader.used()))
311}
312
313/// Decodes a chunk written by [`encode`].
314///
315/// # Errors
316///
317/// If the bytes are truncated, carry an unknown tag, or describe a chunk whose parts disagree.
318pub fn decode(bytes: &[u8]) -> Result<Vec<Vec<u8>>> {
319    Ok(decode_flat(bytes)?.into_values())
320}
321
322/// The size of every candidate that applies, for a report that wants to say what was chosen over
323/// what.
324///
325/// # Errors
326///
327/// As [`encode`].
328pub fn candidate_sizes(values: &[&[u8]]) -> Result<Vec<(Kind, usize)>> {
329    let mut sizes = Vec::new();
330    for kind in candidates(values, 0) {
331        if let Some(bytes) = encode_as(kind, values, 0, &EXHAUSTIVE)? {
332            sizes.push((kind, bytes.len()));
333        }
334    }
335    Ok(sizes)
336}
337
338/// Which candidates [`encode`] would try on this chunk, in the order it tries them.
339///
340/// The chooser is exhaustive, so this is also the list of encodes it pays for to return one of
341/// them. A caller measuring where the encode time goes needs the list separately from the sizes,
342/// because a candidate that is offered and turns out not to apply still costs whatever it spent
343/// finding that out.
344#[must_use]
345pub fn offered(values: &[&[u8]]) -> Vec<Kind> {
346    candidates(values, 0)
347}
348
349/// One candidate on its own, which is what the chooser calls once per entry in [`offered`].
350///
351/// `None` when the encoding does not apply, which is what the chooser treats as a candidate that
352/// did not run rather than as a failure. This is here so that the time the chooser spends can be
353/// attributed to the candidate that spent it, which is the measurement F2 wants before anybody
354/// replaces the exhaustive search with a sampled one. It is not how a writer encodes a chunk:
355/// [`encode`] is, and picking a kind by hand gives up the only thing the chooser is for.
356///
357/// # Errors
358///
359/// As [`encode`].
360pub fn encode_only(kind: Kind, values: &[&[u8]]) -> Result<Option<Vec<u8>>> {
361    encode_as(kind, values, 0, &EXHAUSTIVE)
362}
363
364/// How big one candidate comes out, which is all a sampling chooser needs from it.
365///
366/// The bytes are thrown away, so this says nothing [`encode_only`] does not. It is `pub(crate)` and
367/// separate so that the sampler in [`crate::chooser`] is not handing back buffers it will not read.
368pub(crate) fn size_as(kind: Kind, values: &[&[u8]], depth: u8) -> Result<Option<usize>> {
369    Ok(encode_as(kind, values, depth, &EXHAUSTIVE)?.map(|bytes| bytes.len()))
370}
371
372/// The shape a chunk was encoded as, as a line of text like `DICT(FSST, RLE(...))`.
373///
374/// # Errors
375///
376/// As [`decode`].
377pub fn describe(bytes: &[u8]) -> Result<String> {
378    let mut reader = Reader::new(bytes);
379    describe_chunk(&mut reader)
380}
381
382/// `shape` with one symbol table for the whole column, trained on what reaches FSST in `blocks`.
383///
384/// A settled shape is used for thousands of blocks of one column, and every block that tries FSST
385/// trains its own table. On ClickBench `hits` that was 35 seconds of a 150 second load, most of it
386/// on the literals `FRONT` then `LZ` leaves behind in `URL` and `Referer`, where the table comes
387/// out much the same block after block. So the blocks the shape was settled on are taken down the
388/// shape's levels here, the values that arrive at the FSST level are sampled together, and the
389/// table trained on them is handed to every block through [`Chooser::symbols`].
390///
391/// A shape that ends in `PLAIN` before any FSST level comes back as it was. So does one whose
392/// table comes out empty, which leaves each block to train its own as before.
393#[must_use]
394pub fn with_symbols(shape: Settled, blocks: &[Vec<&[u8]>]) -> Settled {
395    let kinds = shape.strings();
396    let Some(depth) =
397        (0..=kinds.len()).find(|&at| matches!(kinds.get(at), Some(Kind::Fsst) | None))
398    else {
399        return shape;
400    };
401    let leads =
402        kinds[..depth].iter().all(|kind| matches!(kind, Kind::Front | Kind::Lz | Kind::Dict));
403    if !leads || depth > usize::from(MAX_DEPTH) {
404        return shape;
405    }
406    let mut reached: Vec<Vec<u8>> = Vec::new();
407    for block in blocks {
408        let mut values: Vec<Vec<u8>> = block.iter().map(|value| value.to_vec()).collect();
409        for kind in &kinds[..depth] {
410            let refs: Vec<&[u8]> = values.iter().map(Vec::as_slice).collect();
411            values = match kind {
412                Kind::Front => front_code(&refs).1.into_iter().map(<[u8]>::to_vec).collect(),
413                Kind::Dict => dictionary_of(&refs).0.into_iter().map(<[u8]>::to_vec).collect(),
414                _ => {
415                    let joined = refs.concat();
416                    lz::tokens_of(&joined).literals.into_iter().map(<[u8]>::to_vec).collect()
417                }
418            };
419        }
420        reached.extend(values);
421    }
422    let refs: Vec<&[u8]> = reached.iter().map(Vec::as_slice).collect();
423    let table = SymbolTable::train(&sample_of(&refs));
424    if table.is_empty() {
425        return shape;
426    }
427    shape.with_symbols(depth as u8, table)
428}
429
430fn encode_at(values: &[&[u8]], depth: u8, chooser: &dyn Chooser) -> Result<Vec<u8>> {
431    let offered = candidates(values, depth);
432    let mut best: Option<Vec<u8>> = None;
433    for kind in chooser.narrow_strings(values, &offered, depth) {
434        let Some(bytes) = encode_as(kind, values, depth, chooser)? else {
435            continue;
436        };
437        if best.as_ref().is_none_or(|current| bytes.len() < current.len()) {
438            best = Some(bytes);
439        }
440    }
441    best.ok_or_else(|| Error::internal("no string encoding applied to the chunk"))
442}
443
444fn candidates(values: &[&[u8]], depth: u8) -> Vec<Kind> {
445    let mut kinds = vec![Kind::Plain];
446    if values.is_empty() {
447        return kinds;
448    }
449    if values.iter().all(|value| *value == values[0]) {
450        return vec![Kind::Constant];
451    }
452    kinds.push(Kind::Fsst);
453    if depth < MAX_DEPTH && has_duplicates(values) {
454        kinds.push(Kind::Dict);
455    }
456    if depth < MAX_DEPTH && sharing_of(values) >= total_len(values) / SHARE_DIVISOR {
457        kinds.push(Kind::Front);
458    }
459    if depth < MAX_DEPTH && total_len(values) >= LZ_FLOOR {
460        kinds.push(Kind::Lz);
461    }
462    kinds
463}
464
465/// How many bytes each value shares with the value before it, added up.
466///
467/// This is a full pass over the column, and it is here rather than on a sample because it is byte
468/// comparisons that stop at the first difference, which on a column with nothing to share stops
469/// immediately. Against training a symbol table and compressing the whole column, which is what
470/// offering the candidate would cost, it is not worth sampling.
471fn sharing_of(values: &[&[u8]]) -> usize {
472    let mut shared = 0;
473    for pair in values.windows(2) {
474        shared += shared_prefix(pair[0], pair[1]);
475    }
476    shared
477}
478
479/// Every value split into the bytes it shares with the value before it and the bytes it does not.
480///
481/// The suffixes point into the values, so this costs the prefix lengths and nothing else. It is
482/// shared with [`crate::multi`], which front codes a column before compressing it against a symbol
483/// table that belongs to the whole group.
484pub(crate) fn front_code<'a>(values: &[&'a [u8]]) -> (Vec<i64>, Vec<&'a [u8]>) {
485    let mut prefixes = Vec::with_capacity(values.len());
486    let mut suffixes: Vec<&'a [u8]> = Vec::with_capacity(values.len());
487    let mut previous: &[u8] = b"";
488    for value in values {
489        let value: &'a [u8] = value;
490        let shared = shared_prefix(previous, value);
491        prefixes.push(shared as i64);
492        suffixes.push(&value[shared..]);
493        previous = value;
494    }
495    (prefixes, suffixes)
496}
497
498/// The other half. The suffixes are consumed because the values are built out of them.
499///
500/// # Errors
501///
502/// If a prefix is negative or is longer than the value it is a prefix of, which is what a corrupt
503/// or hand written chunk looks like from here.
504pub(crate) fn front_decode(prefixes: &[i64], suffixes: Vec<Vec<u8>>) -> Result<Vec<Vec<u8>>> {
505    let mut values: Vec<Vec<u8>> = Vec::with_capacity(suffixes.len());
506    for (index, suffix) in suffixes.into_iter().enumerate() {
507        let shared = usize::try_from(prefixes[index])
508            .map_err(|_| Error::internal("a negative shared prefix length"))?;
509        let previous: &[u8] = if index == 0 { b"" } else { &values[index - 1] };
510        if shared > previous.len() {
511            return Err(Error::internal(format!(
512                "a value shares {shared} bytes with a value {} bytes long",
513                previous.len()
514            )));
515        }
516        let mut value = Vec::with_capacity(shared + suffix.len());
517        value.extend_from_slice(&previous[..shared]);
518        value.extend_from_slice(&suffix);
519        values.push(value);
520    }
521    Ok(values)
522}
523
524fn shared_prefix(previous: &[u8], value: &[u8]) -> usize {
525    let limit = previous.len().min(value.len());
526    let mut shared = 0;
527    while shared < limit && previous[shared] == value[shared] {
528        shared += 1;
529    }
530    shared
531}
532
533fn total_len(values: &[&[u8]]) -> usize {
534    values.iter().map(|value| value.len()).sum()
535}
536
537fn encode_as(
538    kind: Kind,
539    values: &[&[u8]],
540    depth: u8,
541    chooser: &dyn Chooser,
542) -> Result<Option<Vec<u8>>> {
543    let mut out = vec![kind.tag()];
544    put_u32(&mut out, u32::try_from(values.len()).map_err(|_| too_long(values.len()))?);
545    match kind {
546        Kind::Constant => {
547            let Some(first) = values.first() else {
548                return Ok(None);
549            };
550            if values.iter().any(|value| value != first) {
551                return Ok(None);
552            }
553            put_u32(&mut out, u32::try_from(first.len()).map_err(|_| too_long(first.len()))?);
554            out.extend_from_slice(first);
555        }
556        Kind::Plain => {
557            out.extend_from_slice(&encode_lengths(values, chooser)?);
558            for value in values {
559                out.extend_from_slice(value);
560            }
561        }
562        Kind::Fsst => {
563            let trained;
564            let table = match chooser.symbols(depth) {
565                Some(table) => table,
566                None => {
567                    trained = SymbolTable::train(&sample_of(values));
568                    &trained
569                }
570            };
571            if table.is_empty() {
572                return Ok(None);
573            }
574            let mut compressed = Vec::new();
575            let mut lengths = Vec::with_capacity(values.len());
576            for value in values {
577                let before = compressed.len();
578                table.compress(value, &mut compressed);
579                lengths.push((compressed.len() - before) as i64);
580            }
581            table.serialize(&mut out);
582            out.extend_from_slice(&integer::encode_with(&lengths, chooser)?);
583            out.extend_from_slice(&compressed);
584        }
585        Kind::Dict => {
586            let (entries, codes) = dictionary_of(values);
587            if entries.is_empty() {
588                return Ok(None);
589            }
590            out.extend_from_slice(&encode_at(&entries, depth + 1, chooser)?);
591            out.extend_from_slice(&integer::encode_with(&codes, chooser)?);
592        }
593        Kind::Front => {
594            let (prefixes, suffixes) = front_code(values);
595            out.extend_from_slice(&integer::encode_with(&prefixes, chooser)?);
596            out.extend_from_slice(&encode_at(&suffixes, depth + 1, chooser)?);
597        }
598        Kind::Lz => {
599            let mut joined = Vec::with_capacity(total_len(values));
600            let mut sizes = Vec::with_capacity(values.len());
601            for value in values {
602                joined.extend_from_slice(value);
603                sizes.push(value.len() as i64);
604            }
605            let tokens = lz::tokens_of(&joined);
606            out.extend_from_slice(&integer::encode_with(&sizes, chooser)?);
607            out.extend_from_slice(&integer::encode_with(&tokens.lengths, chooser)?);
608            out.extend_from_slice(&integer::encode_with(&tokens.offsets, chooser)?);
609            out.extend_from_slice(&encode_at(&tokens.literals, depth + 1, chooser)?);
610        }
611    }
612    Ok(Some(out))
613}
614
615fn decode_chunk(reader: &mut Reader<'_>) -> Result<Flat> {
616    let kind = Kind::from_tag(reader.u8()?)?;
617    let count = reader.u32()? as usize;
618    match kind {
619        Kind::Constant => {
620            let len = reader.u32()? as usize;
621            let value = reader.bytes(len)?;
622            let mut flat = Flat::with_capacity(count, len.saturating_mul(count));
623            for _ in 0..count {
624                flat.push(value);
625            }
626            Ok(flat)
627        }
628        Kind::Plain => {
629            let lengths = decode_lengths(reader, count)?;
630            // One copy of the whole payload rather than one a value, which the file already laid
631            // out end to end and which is the layout wanted back.
632            let total = sum_of(&lengths)?;
633            let payload = reader.bytes(total)?;
634            let mut flat = Flat::with_capacity(count, total);
635            flat.bytes.extend_from_slice(payload);
636            let mut at = 0;
637            for length in lengths {
638                at += length;
639                flat.ends.push(at);
640            }
641            Ok(flat)
642        }
643        Kind::Fsst => {
644            let runs = read_compressed(reader, count)?;
645            let mut flat = Flat::with_capacity(count, runs.payload.len());
646            let mut at = 0;
647            for index in 0..count {
648                runs.run_into(index, &mut at, &mut flat.bytes)?;
649                flat.ends.push(flat.bytes.len());
650            }
651            Ok(flat)
652        }
653        Kind::Dict => {
654            let dictionary = decode_chunk(reader)?;
655            let codes = decode_integers(reader)?;
656            if codes.len() != count {
657                return Err(Error::internal(format!(
658                    "a dictionary chunk says it holds {count} values and has {} codes",
659                    codes.len()
660                )));
661            }
662            let mut flat = Flat::with_capacity(count, dictionary.bytes.len());
663            for code in codes {
664                let entry =
665                    usize::try_from(code).ok().and_then(|index| dictionary.get(index)).ok_or_else(
666                        || Error::internal(format!("code {code} is not in the dictionary")),
667                    )?;
668                flat.push(entry);
669            }
670            Ok(flat)
671        }
672        Kind::Front => {
673            let prefixes = decode_integers(reader)?;
674            let suffixes = decode_chunk(reader)?;
675            if prefixes.len() != count || suffixes.len() != count {
676                return Err(Error::internal(format!(
677                    "a front coded chunk says it holds {count} values and has {} prefixes and {} suffixes",
678                    prefixes.len(),
679                    suffixes.len()
680                )));
681            }
682            // The shared prefix is copied out of the buffer being written into, so a value never
683            // has to exist anywhere but where it belongs.
684            let mut flat = Flat::with_capacity(count, suffixes.bytes.len());
685            for (index, prefix) in prefixes.iter().enumerate() {
686                let shared = usize::try_from(*prefix)
687                    .map_err(|_| Error::internal("a negative shared prefix length"))?;
688                let (from, previous) = if index == 0 {
689                    (0, 0)
690                } else {
691                    (flat.start(index - 1), flat.ends[index - 1] - flat.start(index - 1))
692                };
693                if shared > previous {
694                    return Err(Error::internal(format!(
695                        "a value shares {shared} bytes with a value {previous} bytes long"
696                    )));
697                }
698                flat.bytes.extend_from_within(from..from + shared);
699                flat.bytes.extend_from_slice(suffixes.get(index).expect("in range"));
700                flat.ends.push(flat.bytes.len());
701            }
702            Ok(flat)
703        }
704        Kind::Lz => {
705            let sizes = decode_integers(reader)?;
706            let lengths = decode_integers(reader)?;
707            let offsets = decode_integers(reader)?;
708            if sizes.len() != count {
709                return Err(Error::internal(format!(
710                    "a matched chunk says it holds {count} values and has {} lengths",
711                    sizes.len()
712                )));
713            }
714            let mut total = 0usize;
715            let mut widths = Vec::with_capacity(count);
716            for size in sizes {
717                let width = usize::try_from(size)
718                    .map_err(|_| Error::internal("a negative string length"))?;
719                total = total
720                    .checked_add(width)
721                    .ok_or_else(|| Error::internal("a string chunk longer than memory"))?;
722                widths.push(width);
723            }
724            // The copies point back into the bytes already replayed, which is the buffer the values
725            // are going into, so the replay is the decode and there is nothing to cut up after it.
726            let mut flat = Flat::with_capacity(count, total);
727            replay_literals(reader, &lengths, &offsets, total, &mut flat.bytes)?;
728            if flat.bytes.len() != total {
729                return Err(Error::internal(format!(
730                    "a matched chunk rebuilt {} bytes where its lengths add up to {total}",
731                    flat.bytes.len()
732                )));
733            }
734            let mut at = 0;
735            for width in widths {
736                at += width;
737                flat.ends.push(at);
738            }
739            Ok(flat)
740        }
741    }
742}
743
744/// A compressed chunk's symbol table and its runs, left where the file put them.
745///
746/// Reading a compressed chunk into this rather than straight into a buffer is what lets a run be
747/// decompressed where the run belongs. The payload is one slice, the run boundaries come from the
748/// length array, and so asking for a run is a decompress of a subslice and nothing else.
749struct Compressed<'a> {
750    /// The table the runs were compressed against.
751    table: SymbolTable,
752    /// How many compressed bytes each run holds, in order.
753    lengths: Vec<usize>,
754    /// Every run's compressed bytes, end to end.
755    payload: &'a [u8],
756}
757
758impl Compressed<'_> {
759    /// Decompresses run `index` onto the end of `out`, with `at` saying where the run starts.
760    ///
761    /// The caller carries the offset because the runs are asked for in order, and adding a length
762    /// per run is cheaper than the prefix sum the alternative wants.
763    ///
764    /// # Errors
765    ///
766    /// If there is no such run, if it runs off the end of the payload, or if it does not decompress.
767    fn run_into(&self, index: usize, at: &mut usize, out: &mut Vec<u8>) -> Result<()> {
768        self.table.decompress(self.run(index, at)?, out)
769    }
770
771    /// The compressed bytes of run `index`, with `at` saying where the run starts and left where
772    /// the next one does.
773    fn run(&self, index: usize, at: &mut usize) -> Result<&[u8]> {
774        let length = *self
775            .lengths
776            .get(index)
777            .ok_or_else(|| Error::internal(format!("run {index} is not in the chunk")))?;
778        let end = at
779            .checked_add(length)
780            .ok_or_else(|| Error::internal("a compressed chunk longer than memory"))?;
781        let run = self
782            .payload
783            .get(*at..end)
784            .ok_or_else(|| Error::internal("a compressed run is past the end of its chunk"))?;
785        *at = end;
786        Ok(run)
787    }
788}
789
790/// Reads a compressed chunk's table, run lengths and payload without decompressing any of it.
791///
792/// The tag and the count have already been read.
793///
794/// # Errors
795///
796/// If the table does not deserialize, if the length array is not `count` long, or if the lengths
797/// add up to more than the chunk has left.
798fn read_compressed<'a>(reader: &mut Reader<'a>, count: usize) -> Result<Compressed<'a>> {
799    let (table, used) = SymbolTable::deserialize(reader.rest())?;
800    reader.skip(used)?;
801    let lengths = decode_lengths(reader, count)?;
802    // The compressed total is what the payload holds and it is also the only sane guess at the
803    // decompressed one, so it is checked before it is believed.
804    let compressed_len = sum_of(&lengths)?;
805    if compressed_len > reader.remaining() {
806        return Err(Error::internal(format!(
807            "a compressed chunk says it holds {compressed_len} bytes and has {}",
808            reader.remaining()
809        )));
810    }
811    let payload = reader.bytes(compressed_len)?;
812    Ok(Compressed { table, lengths, payload })
813}
814
815/// Replays a matched chunk's tokens, reading the literal runs out of the nested chunk holding them.
816///
817/// The nested chunk is decoded into a buffer and copied out of, the way anything nested is, unless
818/// it is compressed. On the ClickBench `URL` column it always is, and there a block of a thousand
819/// values holds about eight thousand seven hundred literal runs, so that buffer is the whole
820/// block's bytes and copying the runs out of it writes every one of them a second time.
821/// Decompressing a run straight to where it belongs skips the buffer, the length array that would
822/// cut it up, and that second pass over the bytes.
823///
824/// # Errors
825///
826/// Whatever reading the literals or replaying the tokens reports.
827fn replay_literals(
828    reader: &mut Reader<'_>,
829    lengths: &[i64],
830    offsets: &[i64],
831    total: usize,
832    out: &mut Vec<u8>,
833) -> Result<()> {
834    if reader.rest().first() == Some(&Kind::Fsst.tag()) {
835        reader.u8()?;
836        let runs = reader.u32()? as usize;
837        let compressed = read_compressed(reader, runs)?;
838        return replay_in_place(&compressed, lengths, offsets, total, out);
839    }
840    let literals = decode_chunk(reader)?;
841    lz::rebuild_into(&literals, lengths, offsets, out)
842}
843
844/// Room past the end of a replay, for the stores that write whole words past where a value ends.
845///
846/// A symbol is stored as eight bytes and a copy as sixteen at a time, and each is followed by a
847/// step of the cursor to where the bytes it meant end. What lands past that is written over by
848/// whatever comes next, or cut off at the end.
849const REPLAY_SLACK: usize = 16;
850
851/// [`lz::replay`] over compressed literal runs, into a buffer made the length of the output first.
852///
853/// The output length is known before a byte is decoded, because the chunk stores the length of
854/// every value. So the buffer is sized once and written through a cursor, and a symbol or a copy is
855/// a fixed width store rather than a push that checks capacity and moves a length. The copies were
856/// the reason: on ClickBench `URL` a block of a thousand values replays about eight thousand seven
857/// hundred of them, most of them a few tens of bytes, and each one was a call into `memmove`.
858///
859/// # Errors
860///
861/// As [`lz::replay`], and if the tokens build more than `total` bytes.
862fn replay_in_place(
863    compressed: &Compressed<'_>,
864    lengths: &[i64],
865    offsets: &[i64],
866    total: usize,
867    out: &mut Vec<u8>,
868) -> Result<()> {
869    let runs = compressed.lengths.len();
870    if runs != lengths.len() || lengths.len() != offsets.len() {
871        return Err(Error::internal(format!(
872            "a matched chunk has {runs} literal runs, {} lengths and {} offsets",
873            lengths.len(),
874            offsets.len()
875        )));
876    }
877    let base = out.len();
878    let room = total
879        .checked_add(REPLAY_SLACK)
880        .ok_or_else(|| Error::internal("a string chunk longer than memory"))?;
881    out.resize(base + room, 0);
882    let mut read = 0;
883    let mut at = base;
884    for (index, (&length, &offset)) in lengths.iter().zip(offsets).enumerate() {
885        at = compressed.table.decompress_at(compressed.run(index, &mut read)?, out, at)?;
886        let length =
887            usize::try_from(length).map_err(|_| Error::internal("a negative copy length"))?;
888        if length == 0 {
889            continue;
890        }
891        let offset =
892            usize::try_from(offset).map_err(|_| Error::internal("a negative copy offset"))?;
893        at = copy_back(out, base, at, offset, length)?;
894    }
895    if at > base + total {
896        return Err(Error::internal(format!(
897            "a matched chunk rebuilt {} bytes where its lengths add up to {total}",
898            at - base
899        )));
900    }
901    out.truncate(at);
902    Ok(())
903}
904
905/// Copies `length` bytes from `offset` back to `at`, handing back where the copy ends.
906///
907/// Sixteen bytes at a time where the copy starts at least sixteen bytes back, since then no store
908/// reads a byte it has not been given yet, and eight at a time where it starts eight back. Nearer
909/// than that the copy is repeating a short run and goes a byte at a time, the way it always did.
910/// The whole width stores need room past the end of the copy, and a copy near the end of the buffer
911/// that does not have it goes a byte at a time too.
912fn copy_back(
913    out: &mut [u8],
914    base: usize,
915    at: usize,
916    offset: usize,
917    length: usize,
918) -> Result<usize> {
919    if offset == 0 || offset > at - base {
920        return Err(Error::internal(format!(
921            "a copy reaches {offset} bytes back into {} bytes of output",
922            at - base
923        )));
924    }
925    let end = at
926        .checked_add(length)
927        .filter(|&end| end <= out.len())
928        .ok_or_else(|| Error::internal("a matched chunk rebuilds more than its lengths say"))?;
929    let from = at - offset;
930    let wide = end + REPLAY_SLACK <= out.len();
931    if wide && offset >= 16 {
932        let mut step = 0;
933        while step < length {
934            out.copy_within(from + step..from + step + 16, at + step);
935            step += 16;
936        }
937    } else if wide && offset >= 8 {
938        let mut step = 0;
939        while step < length {
940            out.copy_within(from + step..from + step + 8, at + step);
941            step += 8;
942        }
943    } else {
944        for step in 0..length {
945            out[at + step] = out[from + step];
946        }
947    }
948    Ok(end)
949}
950
951fn describe_chunk(reader: &mut Reader<'_>) -> Result<String> {
952    let kind = Kind::from_tag(reader.u8()?)?;
953    let count = reader.u32()? as usize;
954    Ok(match kind {
955        Kind::Constant => {
956            let len = reader.u32()? as usize;
957            reader.bytes(len)?;
958            "CONSTANT".to_string()
959        }
960        Kind::Plain => {
961            let (shape, lengths) = describe_lengths(reader, count)?;
962            reader.skip(lengths.iter().sum())?;
963            format!("PLAIN({shape})")
964        }
965        Kind::Fsst => {
966            let (table, used) = SymbolTable::deserialize(reader.rest())?;
967            reader.skip(used)?;
968            let (shape, lengths) = describe_lengths(reader, count)?;
969            reader.skip(lengths.iter().sum())?;
970            format!("FSST[{}]({shape})", table.len())
971        }
972        Kind::Dict => {
973            let entries = describe_chunk(reader)?;
974            let codes = describe_integers(reader)?;
975            format!("DICT({entries}, {codes})")
976        }
977        Kind::Front => {
978            let prefixes = describe_integers(reader)?;
979            let suffixes = describe_chunk(reader)?;
980            format!("FRONT({prefixes}, {suffixes})")
981        }
982        Kind::Lz => {
983            let sizes = describe_integers(reader)?;
984            let lengths = describe_integers(reader)?;
985            let offsets = describe_integers(reader)?;
986            let literals = describe_chunk(reader)?;
987            format!("LZ({sizes}, {lengths}, {offsets}, {literals})")
988        }
989    })
990}
991
992/// The shape of the length array and the lengths themselves, because a describe has to walk past
993/// the payload to leave the reader where the next chunk starts and the payload size is the sum of
994/// the lengths.
995fn describe_lengths(reader: &mut Reader<'_>, count: usize) -> Result<(String, Vec<usize>)> {
996    let (shape, _) = integer::describe_prefix(reader.rest())?;
997    let lengths = decode_lengths(reader, count)?;
998    Ok((shape, lengths))
999}
1000
1001fn encode_lengths(values: &[&[u8]], chooser: &dyn Chooser) -> Result<Vec<u8>> {
1002    let lengths: Vec<i64> = values.iter().map(|value| value.len() as i64).collect();
1003    integer::encode_with(&lengths, chooser)
1004}
1005
1006fn decode_lengths(reader: &mut Reader<'_>, count: usize) -> Result<Vec<usize>> {
1007    let lengths = decode_integers(reader)?;
1008    if lengths.len() != count {
1009        return Err(Error::internal(format!(
1010            "a string chunk says it holds {count} values and has {} lengths",
1011            lengths.len()
1012        )));
1013    }
1014    lengths
1015        .into_iter()
1016        .map(|length| {
1017            usize::try_from(length).map_err(|_| Error::internal("a negative string length"))
1018        })
1019        .collect()
1020}
1021
1022/// How long the values add up to, refusing a length array that adds up to more than memory.
1023///
1024/// A truncated chunk used to be caught by the read of the value that ran off the end. Reading the
1025/// payload in one go means the total has to be trusted before the read rather than after it, and a
1026/// corrupt length array is the only thing that could overflow it.
1027fn sum_of(lengths: &[usize]) -> Result<usize> {
1028    lengths
1029        .iter()
1030        .try_fold(0usize, |total, length| total.checked_add(*length))
1031        .ok_or_else(|| Error::internal("a string chunk longer than memory"))
1032}
1033
1034/// Reads one nested integer chunk. The integer decoder wants a slice of exactly its own chunk and
1035/// the reader does not know how long that is, so it decodes from the rest of the buffer and is told
1036/// afterwards how much it used.
1037fn decode_integers(reader: &mut Reader<'_>) -> Result<Vec<i64>> {
1038    let (values, used) = integer::decode_prefix(reader.rest())?;
1039    reader.skip(used)?;
1040    Ok(values)
1041}
1042
1043fn describe_integers(reader: &mut Reader<'_>) -> Result<String> {
1044    let (text, used) = integer::describe_prefix(reader.rest())?;
1045    reader.skip(used)?;
1046    Ok(text)
1047}
1048
1049/// A sample of the column spread across the whole of it, taken at random skips rather than at a
1050/// fixed stride.
1051///
1052/// Section 6.3 makes the point about choosing an encoding from a sample and it applies at least as
1053/// much to training a symbol table. Column data is frequently sorted or clustered, so the first
1054/// 64 KB of a URL column is the hosts that sort first and a table trained on it escapes most of the
1055/// rest of the column.
1056///
1057/// The skips are random rather than fixed because a fixed stride aliases. Column data is also
1058/// frequently periodic, and a stride that shares a factor with the period samples one phase of it
1059/// and never sees the others. That is not a hypothetical: the first version of this took every
1060/// `n`th value, and on a test column whose values cycle with a period that the stride happened to
1061/// divide, the table it trained was 3.4 times worse than one trained on the whole column, because
1062/// it learned eight byte symbols that only line up with the phase it saw and had no shorter symbols
1063/// left to fall back on.
1064///
1065/// The generator is a fixed seed xorshift, so the sample is a function of the column and encoding
1066/// the same values twice produces the same bytes.
1067pub(crate) fn sample_of<'a>(values: &[&'a [u8]]) -> Vec<&'a [u8]> {
1068    sample_bytes_of(values, SAMPLE_BYTES)
1069}
1070
1071/// [`sample_of`] with the byte budget spelled out, for a caller training one table over several
1072/// columns that has to split the budget between them.
1073pub(crate) fn sample_bytes_of<'a>(values: &[&'a [u8]], budget: usize) -> Vec<&'a [u8]> {
1074    let budget = budget.max(1);
1075    let total: usize = values.iter().map(|value| value.len()).sum();
1076    if total <= budget {
1077        return values.to_vec();
1078    }
1079    let stride = total.div_ceil(budget).max(1);
1080    let span = (stride * 2 - 1).max(1) as u64;
1081    let mut state = 0x2545_f491_4f6c_dd1du64;
1082    let mut sample = Vec::with_capacity(values.len() / stride + 1);
1083    let mut at = 0usize;
1084    while at < values.len() {
1085        sample.push(values[at]);
1086        state ^= state << 13;
1087        state ^= state >> 7;
1088        state ^= state << 17;
1089        at += 1 + (state % span) as usize;
1090    }
1091    sample
1092}
1093
1094/// The distinct values in sorted order and the code of every value, in one pass over one sort.
1095///
1096/// The dictionary is sorted for the same reason the integer one is: an ordered dictionary turns a
1097/// range predicate into a code range rather than a code set, and front coding over the entries needs
1098/// them sorted anyway.
1099///
1100/// It sorts a permutation of indices rather than the values, which is the whole point. Sorting the
1101/// values means copying every one of them onto the heap first, and the codes then have to be found
1102/// by searching the dictionary back for each value, which is a binary search of string comparisons
1103/// per row. Walking the permutation gives the codes away for free, because the position a value
1104/// sorted to is the position its code was assigned at.
1105fn dictionary_of<'a>(values: &[&'a [u8]]) -> (Vec<&'a [u8]>, Vec<i64>) {
1106    let mut order: Vec<u32> = (0..values.len() as u32).collect();
1107    order.sort_unstable_by(|left, right| values[*left as usize].cmp(values[*right as usize]));
1108    let mut entries: Vec<&'a [u8]> = Vec::new();
1109    let mut codes = vec![0i64; values.len()];
1110    for &index in &order {
1111        let value = values[index as usize];
1112        if entries.last() != Some(&value) {
1113            entries.push(value);
1114        }
1115        codes[index as usize] = (entries.len() - 1) as i64;
1116    }
1117    (entries, codes)
1118}
1119
1120/// Whether any value appears twice, which is the only thing the candidate list wants to know.
1121///
1122/// This used to build the whole sorted dictionary and compare its length against the input, which
1123/// is a copy of the chunk and a sort of it paid on every chunk at every level whether the dictionary
1124/// was ever encoded or not. It is a linear probe over hashes instead: expected O(n), no allocation
1125/// per value, and it stops at the first duplicate it finds, which on a column with any repetition at
1126/// all is immediately.
1127///
1128/// A hash collision is resolved by comparing the bytes, so the answer is exact rather than probable.
1129fn has_duplicates(values: &[&[u8]]) -> bool {
1130    let Some(slots) = values.len().checked_mul(2).map(usize::next_power_of_two) else {
1131        return false;
1132    };
1133    let mask = slots - 1;
1134    let mut table = vec![u32::MAX; slots];
1135    for (index, value) in values.iter().enumerate() {
1136        let mut at = hash_of(value) as usize & mask;
1137        loop {
1138            let held = table[at];
1139            if held == u32::MAX {
1140                table[at] = index as u32;
1141                break;
1142            }
1143            if values[held as usize] == *value {
1144                return true;
1145            }
1146            at = (at + 1) & mask;
1147        }
1148    }
1149    false
1150}
1151
1152/// FNV-1a over the bytes, eight at a time.
1153///
1154/// Good enough for a table that verifies every hit, and it is not part of the format, so nothing
1155/// depends on which hash this is. Eight bytes at a time because a URL column is long values and a
1156/// byte at a time over a hundred bytes of every one of 122,880 rows is the loop this is here to
1157/// avoid.
1158fn hash_of(value: &[u8]) -> u64 {
1159    let mut hash = 0xcbf2_9ce4_8422_2325_u64;
1160    let mut chunks = value.chunks_exact(8);
1161    for chunk in &mut chunks {
1162        let word = u64::from_le_bytes(chunk.try_into().expect("chunks_exact(8) gives eight bytes"));
1163        hash = (hash ^ word).wrapping_mul(0x1_0000_01b3);
1164    }
1165    for byte in chunks.remainder() {
1166        hash = (hash ^ u64::from(*byte)).wrapping_mul(0x1_0000_01b3);
1167    }
1168    (hash ^ (value.len() as u64)).wrapping_mul(0x1_0000_01b3)
1169}
1170
1171fn too_long(len: usize) -> Error {
1172    Error::internal(format!("a string chunk of {len} is longer than the format allows"))
1173}
1174
1175fn put_u32(out: &mut Vec<u8>, value: u32) {
1176    out.extend_from_slice(&value.to_le_bytes());
1177}
1178
1179#[cfg(test)]
1180mod tests {
1181    use super::*;
1182
1183    fn urls(count: usize) -> Vec<Vec<u8>> {
1184        let hosts = ["www.example.com", "shop.example.com", "news.other.example.org"];
1185        let paths = ["/index.html", "/catalog/item", "/search", "/user/profile/settings"];
1186        (0..count)
1187            .map(|index| {
1188                let host = hosts[index % hosts.len()];
1189                let path = paths[(index / 3) % paths.len()];
1190                format!("http://{host}{path}?session={}&ref=google", index * 7).into_bytes()
1191            })
1192            .collect()
1193    }
1194
1195    fn front_lz() -> Settled {
1196        Settled::new(vec![Kind::Front, Kind::Lz], vec![integer::Kind::Packed])
1197    }
1198
1199    /// The point of the table: every block of a column compresses against one table trained once,
1200    /// and what it writes still reads back as the values, including a block the table was not
1201    /// trained on.
1202    #[test]
1203    fn a_block_compressed_against_the_column_table_reads_back() {
1204        let values = urls(4096);
1205        let refs: Vec<&[u8]> = values.iter().map(Vec::as_slice).collect();
1206        let blocks: Vec<Vec<&[u8]>> = refs.chunks(1024).take(2).map(<[&[u8]]>::to_vec).collect();
1207        let shape = with_symbols(front_lz(), &blocks);
1208        assert!(shape.symbols(2).is_some(), "FRONT then LZ leaves FSST the third level");
1209        assert!(shape.symbols(1).is_none(), "and only that one");
1210        for block in refs.chunks(1024) {
1211            let encoded = encode_with(block, &shape).expect("encoded");
1212            assert_eq!(decode(&encoded).expect("decoded"), block.to_vec());
1213        }
1214    }
1215
1216    /// A shape that settles on `PLAIN` never tries FSST, so there is nothing to train.
1217    #[test]
1218    fn a_shape_ending_in_plain_gets_no_table() {
1219        let values = urls(1024);
1220        let refs: Vec<&[u8]> = values.iter().map(Vec::as_slice).collect();
1221        let plain = Settled::new(vec![Kind::Lz, Kind::Plain], vec![integer::Kind::Packed]);
1222        let shape = with_symbols(plain, std::slice::from_ref(&refs));
1223        assert!((0..=MAX_DEPTH).all(|depth| shape.symbols(depth).is_none()));
1224        let fsst = Settled::new(vec![Kind::Fsst], vec![integer::Kind::Packed]);
1225        assert!(with_symbols(fsst, &[refs]).symbols(0).is_some());
1226    }
1227
1228    /// The same values with a scrambled identifier stuck on the front of each, for the tests that
1229    /// need neighbouring values to have nothing in common. Shuffling the order is not enough,
1230    /// because two URLs picked at random still agree on a scheme and often on a host.
1231    fn keyed(values: Vec<Vec<u8>>) -> Vec<Vec<u8>> {
1232        values
1233            .into_iter()
1234            .enumerate()
1235            .map(|(index, value)| {
1236                let key = (index as u64).wrapping_mul(0x9e37_79b9_7f4a_7c15) % 1_000_000_007;
1237                let mut out = format!("{key:010}/").into_bytes();
1238                out.extend_from_slice(&value);
1239                out
1240            })
1241            .collect()
1242    }
1243
1244    fn borrow(values: &[Vec<u8>]) -> Vec<&[u8]> {
1245        values.iter().map(Vec::as_slice).collect()
1246    }
1247
1248    fn round_trip(values: &[Vec<u8>]) -> Vec<u8> {
1249        let borrowed = borrow(values);
1250        let bytes = encode(&borrowed).unwrap();
1251        let back = decode(&bytes).unwrap();
1252        assert_eq!(back, values, "{}", describe(&bytes).unwrap());
1253        check_flat(&bytes, values);
1254        bytes
1255    }
1256
1257    /// The flat form holds the same values and lays them out the way a caller with its own offsets
1258    /// expects. Called from [`round_trip`], so every shape any test in here reaches is checked.
1259    fn check_flat(bytes: &[u8], values: &[Vec<u8>]) {
1260        let flat = decode_flat(bytes).unwrap();
1261        let shape = describe(bytes).unwrap();
1262        assert_eq!(flat.len(), values.len(), "{shape}");
1263        assert_eq!(flat.iter().collect::<Vec<_>>(), borrow(values), "{shape}");
1264        assert_eq!(flat.bytes(), values.concat(), "{shape}");
1265        assert_eq!(flat.get(values.len()), None, "{shape}");
1266    }
1267
1268    fn kind_of(bytes: &[u8]) -> Kind {
1269        Kind::from_tag(bytes[0]).unwrap()
1270    }
1271
1272    #[test]
1273    fn every_shape_decodes_flat_to_what_it_decodes_split() {
1274        // round_trip only sees the shape the chooser picked, which on any one column is one of the
1275        // six. This walks all of them, so PLAIN reading its payload in one go and FRONT copying a
1276        // prefix out of the buffer it is filling are both covered on data they apply to.
1277        let columns =
1278            [urls(600), keyed(urls(600)), vec![b"same".to_vec(); 400], vec![Vec::new(); 7]];
1279        for values in &columns {
1280            let borrowed = borrow(values);
1281            for kind in offered(&borrowed) {
1282                let Some(bytes) = encode_only(kind, &borrowed).unwrap() else {
1283                    continue;
1284                };
1285                assert_eq!(decode(&bytes).unwrap(), *values, "{}", kind.name());
1286                let flat = decode_flat(&bytes).unwrap();
1287                assert_eq!(flat.iter().collect::<Vec<_>>(), borrowed, "{}", kind.name());
1288                assert_eq!(flat.bytes(), values.concat(), "{}", kind.name());
1289            }
1290        }
1291    }
1292
1293    #[test]
1294    fn a_front_coded_chunk_that_shares_more_than_it_has_is_an_error() {
1295        // The prefix chain is the one place the flat decoder reads back out of the buffer it is
1296        // filling, so a prefix longer than the value before it is what would hand back somebody
1297        // else's bytes rather than fail. Built by hand because no encoder produces one.
1298        let suffixes: [&[u8]; 2] = [b"abc", b"x"];
1299        let mut bytes = vec![Kind::Front.tag()];
1300        put_u32(&mut bytes, 2);
1301        bytes.extend_from_slice(&integer::encode(&[0, 9]).unwrap());
1302        bytes.extend_from_slice(&encode_only(Kind::Plain, &suffixes).unwrap().unwrap());
1303        let error = decode_flat(&bytes).expect_err("a nine byte prefix of a three byte value");
1304        assert_eq!(error.message(), "a value shares 9 bytes with a value 3 bytes long");
1305        assert_eq!(decode(&bytes).unwrap_err().message(), error.message());
1306    }
1307
1308    #[test]
1309    fn the_dictionary_is_sorted_and_the_codes_point_back_at_the_values() {
1310        // The two things the dictionary path has to get right, and the reason it is one function
1311        // now rather than a sort followed by a binary search per row.
1312        let values = vec![
1313            b"pear".to_vec(),
1314            b"apple".to_vec(),
1315            b"pear".to_vec(),
1316            b"cherry".to_vec(),
1317            b"apple".to_vec(),
1318        ];
1319        let borrowed = borrow(&values);
1320        let (entries, codes) = dictionary_of(&borrowed);
1321        assert_eq!(entries, vec![b"apple".as_slice(), b"cherry".as_slice(), b"pear".as_slice()]);
1322        assert_eq!(codes, vec![2, 0, 2, 1, 0]);
1323        for (code, value) in codes.iter().zip(&borrowed) {
1324            assert_eq!(entries[*code as usize], *value);
1325        }
1326    }
1327
1328    #[test]
1329    fn a_column_with_nothing_repeated_has_no_duplicates_and_one_with_anything_does() {
1330        let distinct: Vec<Vec<u8>> =
1331            (0..5000).map(|index| format!("value-{index}").into_bytes()).collect();
1332        assert!(!has_duplicates(&borrow(&distinct)));
1333
1334        // One repeat at the far end, so a check that gave up early would miss it.
1335        let mut repeated = distinct.clone();
1336        repeated.push(b"value-0".to_vec());
1337        assert!(has_duplicates(&borrow(&repeated)));
1338
1339        assert!(!has_duplicates(&borrow(&Vec::new())));
1340        assert!(!has_duplicates(&borrow(&[b"one".to_vec()])));
1341        assert!(has_duplicates(&borrow(&vec![b"same".to_vec(); 2])));
1342    }
1343
1344    #[test]
1345    fn long_values_that_differ_only_at_the_end_are_not_confused_for_each_other() {
1346        // The hash is eight bytes at a time and the table verifies every hit, so this is the case
1347        // that says the verify is really there rather than the hash being trusted.
1348        let stem = "http://www.example.com/a/very/long/path/that/goes/on?session=";
1349        let values: Vec<Vec<u8>> =
1350            (0..2000).map(|index| format!("{stem}{index}").into_bytes()).collect();
1351        assert!(!has_duplicates(&borrow(&values)));
1352        let (entries, codes) = dictionary_of(&borrow(&values));
1353        assert_eq!(entries.len(), values.len());
1354        assert_eq!(codes.len(), values.len());
1355    }
1356
1357    #[test]
1358    fn what_the_chooser_returns_is_the_smallest_of_what_it_was_offered() {
1359        // `offered` and `encode_only` are what `cargo xtask encode` splits the chooser's seconds
1360        // with, so they have to describe the chooser that actually runs rather than a second copy
1361        // of its rules that drifts. This is the assertion that keeps the two the same thing: walk
1362        // the list, encode each one alone, and the smallest has to be byte for byte what `encode`
1363        // came back with.
1364        for values in [urls(400), keyed(urls(400)), vec![b"same".to_vec(); 50], Vec::new()] {
1365            let borrowed = borrow(&values);
1366            let chosen = encode(&borrowed).unwrap();
1367            let mut smallest: Option<Vec<u8>> = None;
1368            for kind in offered(&borrowed) {
1369                let Some(bytes) = encode_only(kind, &borrowed).unwrap() else {
1370                    continue;
1371                };
1372                if smallest.as_ref().is_none_or(|best| bytes.len() < best.len()) {
1373                    smallest = Some(bytes);
1374                }
1375            }
1376            assert_eq!(smallest.as_deref(), Some(chosen.as_slice()), "{}", values.len());
1377        }
1378    }
1379
1380    fn raw_size(values: &[Vec<u8>]) -> usize {
1381        values.iter().map(Vec::len).sum::<usize>() + values.len() * 4
1382    }
1383
1384    #[test]
1385    fn a_matched_chunk_replays_literals_whether_or_not_they_are_compressed() {
1386        // The literals of a matched chunk are a chunk of their own, and when that chunk is
1387        // compressed the replay decompresses each run straight into the output instead of into a
1388        // buffer it then copies out of. Both columns here are checked value for value by
1389        // round_trip, so what is left is to show that one of them takes the fused path and the
1390        // other takes the one that decodes the literals first, and that the two agree.
1391        let compressed = describe(&round_trip(&keyed(urls(20_000)))).unwrap();
1392        assert!(compressed.starts_with("LZ(") && compressed.contains(", FSST["), "{compressed}");
1393
1394        let buffered = describe(&round_trip(&keyed(urls(300)))).unwrap();
1395        assert!(buffered.starts_with("LZ(") && buffered.contains(", PLAIN("), "{buffered}");
1396    }
1397
1398    #[test]
1399    fn a_copy_back_writes_what_a_byte_at_a_time_copy_writes_at_every_distance() {
1400        // The wide stores read bytes the same copy wrote a step earlier once the copy is longer
1401        // than its distance, so every distance either side of eight and sixteen is checked against
1402        // the plain loop, at lengths that end short of, on and past a whole store.
1403        let seed: Vec<u8> = (0..40u8).map(|byte| byte.wrapping_mul(37).wrapping_add(11)).collect();
1404        for offset in 1..=seed.len() {
1405            for length in 1..=50 {
1406                let mut wanted = seed.clone();
1407                for _ in 0..length {
1408                    wanted.push(wanted[wanted.len() - offset]);
1409                }
1410                let mut out = seed.clone();
1411                out.resize(seed.len() + length + REPLAY_SLACK, 0);
1412                let end = copy_back(&mut out, 0, seed.len(), offset, length).unwrap();
1413                assert_eq!(&out[..end], wanted.as_slice(), "offset {offset} length {length}");
1414            }
1415        }
1416        let mut short = vec![1, 2, 3, 0];
1417        assert!(copy_back(&mut short, 0, 3, 1, 2).is_err(), "past the end of the buffer");
1418        assert!(copy_back(&mut short, 0, 3, 4, 1).is_err(), "further back than the output");
1419    }
1420
1421    #[test]
1422    fn an_empty_chunk_round_trips() {
1423        let bytes = round_trip(&[]);
1424        assert_eq!(kind_of(&bytes), Kind::Plain);
1425    }
1426
1427    #[test]
1428    fn a_constant_column_costs_what_one_value_costs() {
1429        let values = vec![b"https://www.example.com/".to_vec(); 100_000];
1430        let bytes = round_trip(&values);
1431        assert_eq!(kind_of(&bytes), Kind::Constant);
1432        assert_eq!(bytes.len(), 9 + 24);
1433    }
1434
1435    #[test]
1436    fn a_url_column_of_unique_values_is_matched_rather_than_only_compressed() {
1437        // Every value distinct, so a dictionary is the values plus an index and cannot win, and
1438        // every value starts with an identifier of its own, so neighbours share nothing and front
1439        // coding cannot win either. This used to be the case that fell back to FSST, on the
1440        // reasoning that a symbol table was the only thing that could reach repeated vocabulary
1441        // with no structure around it. That reasoning was wrong and #575 is the measurement: the
1442        // vocabulary repeats at a distance, and a match finder reaches distance where a 255 symbol
1443        // table of at most eight bytes each does not.
1444        let values = keyed(urls(20_000));
1445        let bytes = round_trip(&values);
1446        assert_eq!(kind_of(&bytes), Kind::Lz);
1447
1448        // Against the encoding that used to win, on the same values, so the claim is a comparison
1449        // and not just a label.
1450        let borrowed: Vec<&[u8]> = values.iter().map(Vec::as_slice).collect();
1451        let fsst = encode_as(Kind::Fsst, &borrowed, 0, &EXHAUSTIVE).unwrap().unwrap();
1452        assert!(bytes.len() < fsst.len(), "{} against FSST {}", bytes.len(), fsst.len());
1453
1454        // Eleven bytes of every value are the identifier and a separator and nothing compresses
1455        // them, so the ratio here is lower than the one FSST gets on the URLs on their own.
1456        let ratio = raw_size(&values) as f64 / bytes.len() as f64;
1457        assert!(ratio > 4.0, "{ratio:.2}x");
1458    }
1459
1460    #[test]
1461    fn a_sample_of_a_periodic_column_learns_every_phase_of_it() {
1462        // This column is periodic and its period is what a fixed stride would have divided. The
1463        // sample has to see all of it, because a table trained on one phase learns eight byte
1464        // symbols that only line up with that phase and has nothing shorter to fall back on. The
1465        // measured cost of getting this wrong was 3.4 times the compressed size.
1466        let values = urls(20_000);
1467        let borrowed = borrow(&values);
1468        let sample = sample_of(&borrowed);
1469        let mut phases: Vec<&[u8]> = sample
1470            .iter()
1471            .map(|value| {
1472                let query =
1473                    value.iter().position(|byte| *byte == b'?').expect("every value has a query");
1474                &value[..query]
1475            })
1476            .collect();
1477        phases.sort_unstable();
1478        phases.dedup();
1479        // Three hosts and four paths, and the sample has to contain all twelve of the combinations.
1480        assert_eq!(phases.len(), 12);
1481        let whole = SymbolTable::train(&borrowed);
1482        let sampled = SymbolTable::train(&sample);
1483        let mut on_whole = Vec::new();
1484        let mut on_sample = Vec::new();
1485        for value in &borrowed {
1486            whole.compress(value, &mut on_whole);
1487            sampled.compress(value, &mut on_sample);
1488        }
1489        // Training on a twentieth of the column is allowed to cost something. It is not allowed to
1490        // cost a factor.
1491        assert!(
1492            on_sample.len() < on_whole.len() * 5 / 4,
1493            "{} against {}",
1494            on_sample.len(),
1495            on_whole.len()
1496        );
1497    }
1498
1499    #[test]
1500    fn a_repeating_column_becomes_a_dictionary_of_compressed_entries() {
1501        // The DICT_FSST row of the section 6.2 table, which is not an encoding of its own here: it
1502        // is a dictionary whose entries went back through the chooser. What the entries then get
1503        // is whatever wins on them, and since #575 that is the match finder rather than front
1504        // coding with the leftovers FSST compressed. The point of the test is unchanged: nobody
1505        // named the shape and the chooser arrived at it.
1506        let distinct = urls(500);
1507        let values: Vec<Vec<u8>> =
1508            (0..50_000).map(|index| distinct[index * 7919 % distinct.len()].clone()).collect();
1509        let bytes = round_trip(&values);
1510        assert_eq!(kind_of(&bytes), Kind::Dict);
1511        let shape = describe(&bytes).unwrap();
1512        assert!(shape.starts_with("DICT(LZ("), "{shape}");
1513        let ratio = raw_size(&values) as f64 / bytes.len() as f64;
1514        assert!(ratio > 20.0, "{ratio:.2}x, {shape}");
1515    }
1516
1517    #[test]
1518    fn a_column_of_long_runs_costs_almost_nothing() {
1519        // A dictionary makes the codes an integer chunk, and the integer chunk knows what to do
1520        // with runs, so run length encoding of strings falls out of the recursion.
1521        let distinct = urls(50);
1522        let mut values = Vec::new();
1523        for entry in &distinct {
1524            values.extend(std::iter::repeat_n(entry.clone(), 1000));
1525        }
1526        let bytes = round_trip(&values);
1527        let shape = describe(&bytes).unwrap();
1528        assert!(shape.contains("RLE"), "{shape}");
1529        assert!(bytes.len() < 2000, "{} bytes: {shape}", bytes.len());
1530    }
1531
1532    #[test]
1533    fn incompressible_strings_stay_close_to_their_own_size() {
1534        // The case where nothing works. It has to land on PLAIN or on an FSST that is not much
1535        // worse, rather than on a dictionary of every value in the column.
1536        let mut state = 0x2545_f491_4f6c_dd1du64;
1537        let values: Vec<Vec<u8>> = (0..2000)
1538            .map(|_| {
1539                (0..32)
1540                    .map(|_| {
1541                        state ^= state << 13;
1542                        state ^= state >> 7;
1543                        state ^= state << 17;
1544                        state as u8
1545                    })
1546                    .collect()
1547            })
1548            .collect();
1549        let bytes = round_trip(&values);
1550        assert!(bytes.len() < 2000 * 32 + 3000, "{} bytes", bytes.len());
1551    }
1552
1553    #[test]
1554    fn lengths_are_stored_rather_than_offsets() {
1555        // Every value is 24 bytes, so the lengths are a constant chunk and cost 13 bytes for the
1556        // whole column. Offsets would be 100,000 increasing integers.
1557        let values: Vec<Vec<u8>> =
1558            (0..100_000).map(|index| format!("{index:024}").into_bytes()).collect();
1559        let borrowed = borrow(&values);
1560        let bytes = encode_only(Kind::Plain, &borrowed).unwrap().unwrap();
1561        assert_eq!(bytes.len(), 5 + 13 + 100_000 * 24);
1562    }
1563
1564    #[test]
1565    fn empty_strings_are_values_and_not_nulls() {
1566        let values = vec![Vec::new(), b"a".to_vec(), Vec::new(), b"bb".to_vec()];
1567        round_trip(&values);
1568    }
1569
1570    #[test]
1571    fn a_chunk_with_one_value_round_trips() {
1572        round_trip(&[b"only".to_vec()]);
1573    }
1574
1575    #[test]
1576    fn every_candidate_that_applies_decodes_to_the_input() {
1577        let values = urls(3000);
1578        let borrowed = borrow(&values);
1579        let applicable = candidates(&borrowed, 0);
1580        assert!(applicable.len() >= 2, "{applicable:?}");
1581        for kind in applicable {
1582            let bytes = encode_only(kind, &borrowed).unwrap().unwrap();
1583            assert_eq!(decode(&bytes).unwrap(), values, "{}", kind.name());
1584        }
1585    }
1586
1587    #[test]
1588    fn the_chooser_picks_the_smallest_candidate() {
1589        let values = urls(2000);
1590        let borrowed = borrow(&values);
1591        let chosen = encode(&borrowed).unwrap();
1592        for (_, size) in candidate_sizes(&borrowed).unwrap() {
1593            assert!(chosen.len() <= size);
1594        }
1595    }
1596
1597    #[test]
1598    fn a_truncated_chunk_is_an_error_and_not_a_panic() {
1599        let values = urls(40);
1600        let bytes = encode(&borrow(&values)).unwrap();
1601        for len in 0..bytes.len() {
1602            assert!(decode(&bytes[..len]).is_err(), "{len} bytes decoded");
1603        }
1604    }
1605
1606    #[test]
1607    fn trailing_bytes_are_an_error() {
1608        let mut bytes = encode(&borrow(&urls(10))).unwrap();
1609        bytes.push(0);
1610        let error = decode(&bytes).unwrap_err();
1611        assert!(error.message().contains("left over"), "{error}");
1612    }
1613
1614    #[test]
1615    fn an_unknown_tag_is_an_error() {
1616        let error = decode(&[99, 0, 0, 0, 0]).unwrap_err();
1617        assert!(error.message().contains("unknown string encoding tag"), "{error}");
1618    }
1619
1620    #[test]
1621    fn a_dictionary_code_outside_the_dictionary_is_an_error() {
1622        let mut bytes = vec![Kind::Dict.tag()];
1623        put_u32(&mut bytes, 1);
1624        bytes.extend_from_slice(&encode(&[b"one".as_slice()]).unwrap());
1625        bytes.extend_from_slice(&integer::encode(&[9]).unwrap());
1626        let error = decode(&bytes).unwrap_err();
1627        assert!(error.message().contains("not in the dictionary"), "{error}");
1628    }
1629
1630    #[test]
1631    fn a_sorted_column_of_urls_is_front_coded() {
1632        // The M1 finding, in a test. Sorted URLs share a host and most of a path with the URL next
1633        // to them, FSST cannot reach those bytes because it compresses each value on its own, and
1634        // front coding is the shape that reaches them.
1635        let mut values = urls(20_000);
1636        values.sort();
1637        let bytes = round_trip(&values);
1638        assert_eq!(kind_of(&bytes), Kind::Front);
1639        let shape = describe(&bytes).unwrap();
1640        let mut plain = Vec::new();
1641        let borrowed = borrow(&values);
1642        for (kind, size) in candidate_sizes(&borrowed).unwrap() {
1643            if kind == Kind::Fsst {
1644                plain.push(size);
1645            }
1646        }
1647        let fsst = plain[0];
1648        assert!(bytes.len() * 2 < fsst, "{} against FSST {fsst}: {shape}", bytes.len());
1649    }
1650
1651    #[test]
1652    fn a_column_with_nothing_to_share_is_not_offered_front_coding() {
1653        // The candidate costs an encode of the whole column, so a column whose neighbours have
1654        // nothing in common must not be paying for it.
1655        let mut state = 0x9e37_79b9_7f4a_7c15u64;
1656        let values: Vec<Vec<u8>> = (0..2000)
1657            .map(|_| {
1658                (0..24)
1659                    .map(|_| {
1660                        state ^= state << 13;
1661                        state ^= state >> 7;
1662                        state ^= state << 17;
1663                        (state % 251) as u8
1664                    })
1665                    .collect()
1666            })
1667            .collect();
1668        let borrowed = borrow(&values);
1669        assert!(!candidates(&borrowed, 0).contains(&Kind::Front));
1670    }
1671
1672    #[test]
1673    fn a_prefix_longer_than_the_value_before_it_is_an_error() {
1674        let mut bytes = vec![Kind::Front.tag()];
1675        put_u32(&mut bytes, 2);
1676        bytes.extend_from_slice(&integer::encode(&[0, 9]).unwrap());
1677        bytes.extend_from_slice(&encode(&[b"one".as_slice(), b"two".as_slice()]).unwrap());
1678        let error = decode(&bytes).unwrap_err();
1679        assert!(error.message().contains("shares 9 bytes"), "{error}");
1680    }
1681
1682    #[test]
1683    fn a_negative_prefix_is_an_error() {
1684        let mut bytes = vec![Kind::Front.tag()];
1685        put_u32(&mut bytes, 1);
1686        bytes.extend_from_slice(&integer::encode(&[-1]).unwrap());
1687        bytes.extend_from_slice(&encode(&[b"one".as_slice()]).unwrap());
1688        let error = decode(&bytes).unwrap_err();
1689        assert!(error.message().contains("negative shared prefix"), "{error}");
1690    }
1691
1692    #[test]
1693    fn a_negative_length_is_an_error() {
1694        let mut bytes = vec![Kind::Plain.tag()];
1695        put_u32(&mut bytes, 1);
1696        bytes.extend_from_slice(&integer::encode(&[-1]).unwrap());
1697        let error = decode(&bytes).unwrap_err();
1698        assert!(error.message().contains("negative string length"), "{error}");
1699    }
1700
1701    #[test]
1702    fn the_sample_is_spread_across_the_chunk_and_not_taken_from_the_front() {
1703        // A sorted column whose first 64 KB says nothing about the rest of it. If the sample were
1704        // the front, the table would learn `aaaa` and escape every `zzzz`.
1705        let mut values: Vec<Vec<u8>> = Vec::new();
1706        for index in 0..20_000 {
1707            let head = if index < 10_000 { "aaaaaaaaaaaaaaaa" } else { "zzzzzzzzzzzzzzzz" };
1708            values.push(format!("{head}/{index:08}").into_bytes());
1709        }
1710        let borrowed = borrow(&values);
1711        let sample = sample_of(&borrowed);
1712        let first_half = sample.iter().filter(|value| value.starts_with(b"aaaa")).count();
1713        let second_half = sample.len() - first_half;
1714        assert!(first_half > 0 && second_half > 0, "{first_half} and {second_half}");
1715        let bytes = round_trip(&values);
1716        let ratio = raw_size(&values) as f64 / bytes.len() as f64;
1717        assert!(ratio > 4.0, "{ratio:.2}x");
1718    }
1719}