yo-kv 0.3.8

//! One database, and the parts of it that are not about any particular type.
//!
//! This is the `dict` a Redis `SELECT` picks between, and one of these is what a
//! shard owns. It was called `Strings` while strings were the only thing in it,
//! which was accurate for M2 and stopped being accurate the moment a set needed
//! somewhere to live.
//!
//! The commands hang off this as separate `impl` blocks, one file per type, so
//! that `SET` lives in [`strings`](crate::strings) next to the other twenty five
//! string commands rather than in a file that is the whole of Redis. They are
//! methods on the keyspace and not on some per type object because a key belongs
//! to the database and not to a type: `DEL` does not care what it is deleting,
//! and `SADD` against a string has to be able to see that it is a string.
//!
//! # Not Sync
//!
//! Like everything that hangs off a shard. One of these belongs to one thread and
//! is reached by sending that thread a command, which is Y1, and it is why
//! nothing here takes a lock or an atomic.

use std::sync::atomic::{AtomicU64, Ordering};

use yo_common::{Addr, Code, Error, Result, Rng, bytes_eq};
use yo_index::RawMap;

use crate::Clock;
use crate::access::{Access, Lfu, Policy};
use crate::array::Array;
use crate::evict;
use crate::hash::{self, Hash};
use crate::list::{self, List};
use crate::set::{self, Set};
use crate::slab::{Bytes, Slab};
use crate::ttl::{self, Applied, Ask, Cond};
use crate::value::{self, Kind};
use crate::zset::{self, Zset};

/// Every collection already answered this question, and this is the answer said
/// once more in a shape the slab can ask for without knowing what it is holding.
///
/// Here rather than in the five type files because it is one fact about the
/// keyspace and not five facts about five types, and because a reader looking
/// for how the memory total is kept should find it next to the slabs it counts.
macro_rules! bytes {
    ($($t:ty),*) => { $(impl Bytes for $t {
        #[inline]
        fn memory_bytes(&self) -> usize {
            <$t>::memory_bytes(self)
        }
    })* };
}
bytes!(Set, Hash, List, Zset, Array);

/// One database: every key, whatever type it holds.
pub struct Keyspace {
    pub(crate) map: RawMap,
    pub(crate) clock: Clock,
    /// Keys that were found dead on the way to answering something else.
    pub(crate) expired: u64,
    /// Keys thrown away to make room, which is a different number entirely.
    ///
    /// Redis keeps `expired_keys` and `evicted_keys` apart in `INFO` and the
    /// distinction is the one people watch: expiry is the client getting what it
    /// asked for, and eviction is the server deciding it cannot keep a promise
    /// nobody asked it to break.
    pub(crate) evicted: u64,
    /// Every set in this database, addressed by the number in its record.
    pub(crate) sets: Slab<Set>,
    /// Every hash in this database, addressed the same way.
    ///
    /// A slab per type rather than one slab of an enum, so that a record's four
    /// bytes index a `Hash` directly and reaching one is a load and not a load
    /// followed by a discriminant check. The type tag in the record already
    /// says which slab to look in, so the discriminant would be a second copy
    /// of a fact the record has.
    pub(crate) hashes: Slab<Hash>,
    /// Every list in this database, addressed the same way.
    pub(crate) lists: Slab<List>,
    /// Every sorted set in this database, addressed the same way.
    pub(crate) zsets: Slab<Zset>,
    /// Every sparse array in this database, addressed the same way.
    pub(crate) arrays: Slab<Array>,
    /// How many keys hold something that is not a string.
    ///
    /// This exists so that a database of nothing but strings, which is every
    /// benchmark today and most of what `SET` sees, can skip the body check in
    /// [`Keyspace::free_body`] on one predictable branch against a field that is
    /// already hot, rather than paying a second lookup per write forever.
    pub(crate) bodies: usize,
    /// Where a set changes representation.
    pub(crate) limits: set::Limits,
    /// Where a hash changes representation.
    pub(crate) hash_limits: hash::Limits,
    /// Where a list changes representation.
    pub(crate) list_limits: list::Limits,
    /// Where a sorted set changes representation.
    pub(crate) zset_limits: zset::Limits,
    /// What this database would evict, and therefore what a read writes back.
    ///
    /// One server wide setting in Redis, carried per database here for the same
    /// reason the size ladder is: a `Keyspace` is reached without a server and
    /// has to be able to answer on its own. `CONFIG SET maxmemory-policy` writes
    /// it to all of them.
    pub(crate) policy: Policy,
    /// The two numbers the LFU counter moves by, which are `CONFIG` values.
    pub(crate) lfu: Lfu,
    /// How many keys a round of eviction sampling looks at.
    ///
    /// `maxmemory-samples`, carried per database for the same reason the policy
    /// is. See [`evict::SAMPLES`] for why the default is five.
    pub(crate) samples: usize,
    /// Where `SPOP` and `SRANDMEMBER` draw from.
    pub(crate) rng: Rng,
    /// The last collection key that was resolved, for the command behind it.
    memo: Memo,
    /// One buffer for the commands that have to hold an element while the
    /// structure it came out of is being written.
    ///
    /// [`Keyspace::lmove`] is the reason this is here: it takes an element out
    /// of one list and puts it into another, so there is a moment where the
    /// bytes belong to nothing, and the borrow it would need to avoid that is a
    /// borrow of two lists at once when the two lists may be the same one. A
    /// `Vec` per call is the obvious way to cover that moment and it is a malloc
    /// and a free on a command that a queue sends millions of. This is the same
    /// `Vec` every time, cleared rather than freed, so the steady state is no
    /// allocator call at all.
    ///
    /// [`Keyspace::append`], [`Keyspace::setrange`] and the string arm of
    /// [`Keyspace::set_expiry`] use it for the same shape of problem: each of
    /// them has to hold the old value while it writes the new record, and each
    /// of them was doing that with a fresh `Vec` of the whole value. They cannot
    /// overlap, because each one puts the buffer back before it returns and one
    /// command runs at a time.
    ///
    /// It lives on the database and not on the caller because the callers are
    /// wire handlers that are handed a `&mut Keyspace` and nothing else.
    ///
    /// It starts at [`SCRATCH`] bytes rather than empty. An empty one grows on
    /// the first command that uses it, and that growth is a real allocation on a
    /// command path even though it happens once. Buying it here, where nobody is
    /// waiting, makes the rule Y7 enforces true without an exception written for
    /// it. A value larger than that still grows it, and that one is allocation
    /// proportional to what the caller sent rather than overhead per command.
    pub(crate) scratch: Vec<u8>,

    /// The same idea for indices rather than bytes.
    ///
    /// `ZRANDMEMBER` with a positive count under the size of the set does a
    /// partial Fisher-Yates, and that needs the permutation somewhere while it
    /// draws from it. One buffer, cleared and refilled, rather than one `Vec`
    /// per call, because sampling is a thing callers do in a loop.
    ///
    /// It does not start at a capacity, unlike [`Keyspace::scratch`]. There is
    /// no size to guess: the buffer has to be as long as the set, so the first
    /// call on a set larger than anything seen before grows it whatever it was
    /// given to start with. That growth is proportional to the data rather than
    /// per command.
    pub(crate) rows: Vec<usize>,

    /// The tables set algebra fills in, kept rather than built per call.
    ///
    /// Same idea again, one level up: a union walks everything into a hash
    /// table and lets the table be the duplicate check, and building that table
    /// was the largest single allocation left on any command path. See
    /// [`setops::Scratch`], which is where the two tables and the argument for
    /// them live.
    ///
    /// It is one table per database and not one per command, so a database that
    /// has answered a union over a million members holds a million member table
    /// until it answers a smaller one. That is the trade and it is the right way
    /// round: the same database had to build that table anyway, and the version
    /// that threw it away afterwards built it again on the next call.
    pub(crate) setops: crate::setops::Scratch,
}

/// How big [`Keyspace::scratch`] starts.
///
/// A kibibyte, which covers a value of any ordinary size and costs one
/// allocation per database. The number is not tuned and does not need to be: too
/// small only means the buffer grows once more on some later command, and too
/// large only means a kibibyte nobody used.
const SCRATCH: usize = 1024;

/// How many segments one call to [`Keyspace::victim`] will draw from.
///
/// A round is a whole segment, and a segment is sixty four buckets of seven
/// entries each before its overflow chains are counted, so one round almost
/// always answers. The retries are for the case where a round came back with
/// nothing usable, which happens when the segment it drew was empty or when a
/// `volatile` policy filtered out everything in it.
///
/// Four rather than more, because the case that would want more is the case no
/// number of rounds can fix. A `volatile` policy on a database where nothing has
/// a deadline has no eligible key anywhere, and every round of it is wasted work
/// on a path where a client is waiting. Redis does not have this problem because
/// it keeps a second dictionary of just the keys with deadlines and samples that
/// one directly, which is the real answer here too and is not this change.
const ROUNDS: usize = 4;

/// Where the last collection key resolved to, if it still resolves there.
///
/// Y13 says a batch of `SADD` on one key should be one table growth check, and
/// the same argument applies a step earlier: it should be one resolve. A
/// resolve is a hash, a bucket walk and a record read, and on a hot key every
/// command in the batch was paying for all three to be told the same answer the
/// command in front of it got.
///
/// One entry and not a cache, because one entry is the shape of the problem.
/// Single key `SADD` is the case with no spread to exploit, so the only reuse
/// there is to find is the command immediately before, and a bigger structure
/// would cost a lookup to avoid a lookup.
///
/// It holds a slot rather than reaching the body through an address, because a
/// slot is an index into the slab for its type and stays right for as long as
/// the key is there. That is also why nothing here memoizes a string: a string
/// lives in the record itself and moves when the record does.
///
/// It does carry the record's address alongside, and that is safe for a narrower
/// reason than the slot is. An address is only good until the next write, and
/// this whole memo is thrown away by the next write, so inside the window where
/// the memo answers at all the address is exactly as valid as the slot. What it
/// buys is the eviction stamp: a memo hit skips the probe, and without an address
/// the stamp would have to put the probe back and there would be no memo left.
///
/// What it is worth is measured rather than argued about, by the pair of rows
/// `engine/sadd` and `engine/sadd-alternating` in `yo-resp`'s `engine` bench.
/// The second one alternates between two keys, which defeats this on every
/// command and leaves both keys as warm in the cache as the one key was, so the
/// difference between the rows is close to this and nothing else. On an Apple M4
/// it is about nineteen nanoseconds a command, which is 1.25x at pipeline 64.
struct Memo {
    /// What the map's write counter said when this was taken.
    writes: u64,
    /// Whether there is anything here. Separate from the length because the
    /// empty key is a key, and `SADD "" m` is a command Redis accepts.
    live: bool,
    /// The type the key held, so a hit can still answer `WRONGTYPE`.
    kind: Kind,
    /// Where the body is in the slab for `kind`.
    slot: u32,
    /// Where the record is, for the stamp a hit still owes.
    addr: Addr,
    /// How much of `key` is the key.
    len: u8,
    key: [u8; Memo::MAX],
}

impl Memo {
    /// The longest key worth remembering.
    ///
    /// Thirty two bytes is half a cache line and covers every hot key anyone
    /// writes down, including the `myset:{tag}` the generators send. A longer
    /// key is not memoized rather than heap allocated, because the whole point
    /// of this is to not touch memory it does not have to.
    const MAX: usize = 32;

    const fn empty() -> Memo {
        Memo {
            writes: 0,
            live: false,
            kind: Kind::String,
            slot: 0,
            addr: Addr::NONE,
            len: 0,
            key: [0; Memo::MAX],
        }
    }

    /// What `key` resolved to last time, if that answer still stands.
    ///
    /// `writes` is the map's counter now. Any write at all since this was taken
    /// and the answer is thrown away, which is stricter than it has to be and is
    /// the version that cannot be wrong.
    #[inline]
    fn get(&self, writes: u64, key: &[u8]) -> Option<(Kind, u32, Addr)> {
        if !self.live || self.writes != writes || key.len() != self.len as usize {
            return None;
        }
        // `bytes_eq` and not `==`, which is a call into the platform's `memcmp`
        // for a key of a length the compiler cannot see. This is the one
        // comparison the hot key path always does, and on a profile of `SADD`
        // it was most of what the lookup cost.
        bytes_eq(&self.key[..key.len()], key).then_some((self.kind, self.slot, self.addr))
    }

    /// Remember that `key` is at `slot`, in the record at `addr`.
    #[inline]
    fn put(&mut self, writes: u64, key: &[u8], kind: Kind, slot: u32, addr: Addr) {
        if key.len() > Memo::MAX {
            self.live = false;
            return;
        }
        self.writes = writes;
        self.live = true;
        self.kind = kind;
        self.slot = slot;
        self.addr = addr;
        self.len = key.len() as u8;
        self.key[..key.len()].copy_from_slice(key);
    }
}

/// How many databases this process has made.
///
/// Mixed into a new database's seed so that the eight shards a server starts in
/// the same millisecond do not all draw the same members in the same order. It
/// is the only atomic in this file and it is touched once per database rather
/// than once per command, so it is not on any path Y1 cares about.
static MADE: AtomicU64 = AtomicU64::new(0);

impl Keyspace {
    /// An empty database on the system clock.
    #[must_use]
    pub fn new() -> Keyspace {
        Keyspace::with_clock(Clock::system())
    }

    /// An empty database on a clock of the caller's choosing.
    #[must_use]
    pub fn with_clock(clock: Clock) -> Keyspace {
        let made = MADE.fetch_add(1, Ordering::Relaxed);
        Keyspace {
            map: RawMap::new(),
            clock,
            expired: 0,
            evicted: 0,
            sets: Slab::new(),
            hashes: Slab::new(),
            lists: Slab::new(),
            zsets: Slab::new(),
            arrays: Slab::new(),
            bodies: 0,
            limits: set::Limits::DEFAULT,
            hash_limits: hash::Limits::DEFAULT,
            list_limits: list::Limits::default(),
            zset_limits: zset::Limits::DEFAULT,
            policy: Policy::default(),
            lfu: Lfu::DEFAULT,
            samples: evict::SAMPLES,
            rng: Rng::new(clock.now_ms() ^ made.wrapping_mul(0x9e37_79b9_7f4a_7c15)),
            memo: Memo::empty(),
            scratch: Vec::with_capacity(SCRATCH),
            rows: Vec::new(),
            setops: crate::setops::Scratch::new(),
        }
    }

    /// Pin what `SPOP` and `SRANDMEMBER` draw.
    ///
    /// A database seeds itself from the clock and a counter, which is what a
    /// server wants and what a test cannot assert against. Every test in this
    /// crate that cares which member comes back calls this first, the same way
    /// every expiry test drives a fixed clock, and for the same reason: the one
    /// input that makes a result unrepeatable is better handed in than reached
    /// for.
    ///
    /// It is public because reproducing a bug report is the same problem. A
    /// seed printed in a crash report is worth having somewhere to put.
    #[inline]
    pub const fn seed(&mut self, seed: u64) {
        self.rng = Rng::new(seed);
    }

    /// What this database would evict, which is `CONFIG GET maxmemory-policy`.
    #[inline]
    #[must_use]
    pub const fn policy(&self) -> Policy {
        self.policy
    }

    /// Change what this database would evict.
    ///
    /// Every key already stored keeps whatever is in its access field, which is
    /// why Redis warns on `OBJECT FREQ` that switching at runtime takes time to
    /// adjust. Under the new policy those bits mean something else, and the only
    /// honest thing to do about it is to let them be corrected by use. A key
    /// nobody has touched since the switch reads as freshly used rather than as
    /// stale, which is the safe direction: the other one evicts the working set
    /// on the first pass after an operator changes a setting.
    #[inline]
    pub const fn set_policy(&mut self, policy: Policy) {
        self.policy = policy;
    }

    /// The two numbers the LFU counter moves by, which are two `CONFIG` values.
    #[inline]
    #[must_use]
    pub const fn lfu(&self) -> Lfu {
        self.lfu
    }

    /// Change how fast the LFU counter climbs and decays.
    #[inline]
    pub const fn set_lfu(&mut self, lfu: Lfu) {
        self.lfu = lfu;
    }

    /// Seconds since `key` was last used, which is `OBJECT IDLETIME`.
    ///
    /// `None` for a key that is not there. A key that has never been stamped
    /// reads as zero rather than as ancient, which is what
    /// [`Access::is_unset`] is for.
    ///
    /// This does not count as a use. Redis looks the key up with its no touch
    /// flag here, and it has to: a diagnostic that resets the number it reports
    /// would answer zero every time it was asked.
    pub fn idle_secs(&mut self, key: &[u8]) -> Option<u64> {
        let addr = self.live_rec_untouched(key)?;
        let now = self.clock.now_ms();
        Some(self.access_at(addr).idle_secs(now))
    }

    /// How often `key` is used, which is `OBJECT FREQ`.
    ///
    /// The eight bit counter, decayed to now, on the same terms as
    /// [`Keyspace::idle_secs`]: `None` for a key that is not there, and asking
    /// is not using.
    ///
    /// The caller is the one that has to check the policy first. This reports
    /// what the bits say, and under a policy that is not LFU they say something
    /// else, which is a refusal on the wire rather than a number.
    pub fn freq(&mut self, key: &[u8]) -> Option<u8> {
        let addr = self.live_rec_untouched(key)?;
        let (now, lfu) = (self.clock.now_ms(), self.lfu);
        Some(self.access_at(addr).freq(now, lfu))
    }

    /// Write a record under `key`, with the access field the policy wants on it.
    ///
    /// Every record this crate writes goes through here, which is the point of
    /// it. A record is written fresh whenever a key is created and whenever a
    /// string's value changes, and a fresh record starts with the blank field
    /// [`value::write_record`] leaves behind. Blank reads as freshly used, which
    /// is right at the moment of writing and wrong a minute later, so something
    /// has to stamp it and this is the only place that knows the clock.
    ///
    /// Redis stamps at the same moment, in `createObject`, and for the same
    /// reason.
    pub(crate) fn write_rec(
        &mut self,
        key: &[u8],
        len: usize,
        fill: impl FnOnce(&mut [u8]),
    ) -> Option<usize> {
        let a = self.access_for_write(key);
        self.map.set_with(key, len, |out| {
            fill(out);
            value::set_access(out, a);
        })
    }

    /// What the access field of a record about to be written should say.
    ///
    /// Under the eight policies that read the field as a clock this is the time,
    /// with no probe and no thought: writing a key is using it, and under the LRM
    /// pair writing it is the only thing that counts as using it.
    ///
    /// Under LFU it is the counter that is already there, carried across the
    /// rewrite unchanged. Unchanged rather than incremented, because the lookup
    /// that resolved the key for this write already counted the access, and
    /// counting it twice would rank a key that is written more highly than a key
    /// that is read the same number of times. A key that is not there yet starts
    /// at [`crate::access::LFU_INIT`], which is where Redis starts a new object.
    ///
    /// The probe is the reason this is written as two cases rather than one. It
    /// is paid only under an LFU policy, so the default policy and every other
    /// one write a record for exactly what it cost before.
    fn access_for_write(&mut self, key: &[u8]) -> Access {
        let now = self.clock.now_ms();
        if !self.policy.is_lfu() {
            return Access::lru(now);
        }
        match self.map.get(key).and_then(value::access) {
            Some(a) if !a.is_unset() => a,
            _ => Access::lfu(now),
        }
    }

    /// The access field of the record at `addr`, or the unset one for a record
    /// written before the field existed.
    #[inline]
    fn access_at(&self, addr: Addr) -> Access {
        value::access(self.map.value_at(addr)).unwrap_or_default()
    }

    /// Write the access field back to the record at `addr`.
    ///
    /// The whole reason the field exists, and it runs on nearly every command,
    /// so what it does is a load, an arithmetic step and a three byte store into
    /// a cache line the caller has just read. It does not count as a write to the
    /// map, because nothing moves and counting it would throw the [`Memo`] away
    /// once per command. See [`RawMap::value_at_mut`].
    ///
    /// The LFU arm reads before it writes, because the counter it produces is a
    /// function of the counter that is there. The clock arm does not, because the
    /// time is the time whatever the record used to say.
    #[inline]
    fn stamp(&mut self, addr: Addr) {
        let now = self.clock.now_ms();
        if self.policy.is_lfu() {
            let (lfu, current) = (self.lfu, self.access_at(addr));
            let next = current.touched(now, lfu, &mut self.rng);
            value::set_access(self.map.value_at_mut(addr), next);
        } else {
            value::set_access(self.map.value_at_mut(addr), Access::lru(now));
        }
    }

    /// Where a set changes representation, which is three `CONFIG` values.
    #[inline]
    pub const fn limits(&self) -> &set::Limits {
        &self.limits
    }

    /// Change where a set changes representation.
    ///
    /// Moving these does not rewrite the sets that already exist, which is what
    /// Redis does too: `CONFIG SET set-max-listpack-entries 0` leaves every
    /// listpack alone and only decides what the next `SADD` builds.
    #[inline]
    pub const fn set_limits(&mut self, limits: set::Limits) {
        self.limits = limits;
    }

    /// Where a hash changes representation, which is two `CONFIG` values.
    #[inline]
    pub const fn hash_limits(&self) -> &hash::Limits {
        &self.hash_limits
    }

    /// Change where a hash changes representation.
    ///
    /// Same rule as the set: moving these leaves every hash that already exists
    /// exactly as it is, and only decides what the next `HSET` builds.
    #[inline]
    pub const fn set_hash_limits(&mut self, limits: hash::Limits) {
        self.hash_limits = limits;
    }

    /// Where a list changes representation, which is one `CONFIG` value.
    #[inline]
    pub const fn list_limits(&self) -> &list::Limits {
        &self.list_limits
    }

    /// Change where a list changes representation.
    ///
    /// Same rule again: this decides what the next `LPUSH` builds and leaves
    /// every list that already exists alone. `list-max-listpack-size` is one
    /// number rather than two, and [`list::Limits::of`] is what turns it into
    /// the pair this holds.
    #[inline]
    pub const fn set_list_limits(&mut self, limits: list::Limits) {
        self.list_limits = limits;
    }

    /// Where a sorted set changes representation, which is two `CONFIG` values.
    #[inline]
    pub const fn zset_limits(&self) -> &zset::Limits {
        &self.zset_limits
    }

    /// Change where a sorted set changes representation.
    ///
    /// Same rule as the other three: this decides what the next `ZADD` builds
    /// and leaves every sorted set that already exists exactly as it is.
    #[inline]
    pub const fn set_zset_limits(&mut self, limits: zset::Limits) {
        self.zset_limits = limits;
    }

    /// The clock expiry compares against.
    #[inline]
    pub const fn clock(&self) -> &Clock {
        &self.clock
    }

    /// The clock, to refresh once per turn of the loop.
    #[inline]
    pub const fn clock_mut(&mut self) -> &mut Clock {
        &mut self.clock
    }

    /// The map underneath, for statistics and for compaction.
    #[inline]
    pub const fn map(&self) -> &RawMap {
        &self.map
    }

    /// How many keys are stored, including any that are dead and not yet
    /// noticed. This is Redis's `DBSIZE`, which counts the same way.
    #[inline]
    pub fn len(&self) -> usize {
        self.map.len()
    }

    /// Whether anything is stored.
    #[inline]
    pub fn is_empty(&self) -> bool {
        self.map.is_empty()
    }

    /// What `key` holds, or `None` if there is nothing under it.
    ///
    /// This is `TYPE`. A key past its deadline is reaped first, so a dead key
    /// answers `None` and not the type it used to be.
    ///
    /// One lookup, because the tag and the deadline are both in the record the
    /// lookup returned. Reading the kind out before the reap rather than after
    /// is what keeps it to one.
    ///
    /// It does not go through the lookup that stamps, and so it leaves the
    /// eviction clock where it was, which is right and is worth saying rather than
    /// leaving to be inferred from the shape of the code. `TYPE` is one of the
    /// commands Redis looks up with its no touch flag, along with the `OBJECT`
    /// subcommands underneath this, which read through `reap` for the same
    /// reason.
    pub fn kind_of(&mut self, key: &[u8]) -> Option<Kind> {
        let now = self.clock.now_ms();
        let (kind, dead) = self
            .map
            .get(key)
            .map(|rec| (value::kind(rec), value::is_expired(rec, now)))?;
        if dead {
            self.drop_key(key);
            self.expired += 1;
            return None;
        }
        Some(kind)
    }

    /// How a set is represented, or `None` if `key` is not a set.
    ///
    /// This follows the slot and asks the body rather than reading the record,
    /// because the record only holds a number. Putting a copy of the
    /// representation in the record's two spare encoding bits would mean
    /// rewriting the record every time a set was promoted, for the sake of a
    /// command nobody calls in a loop, and would leave two places able to
    /// disagree about the same fact.
    pub fn set_encoding(&mut self, key: &[u8]) -> Option<set::Encoding> {
        self.reap(key);
        let rec = self.map.get(key)?;
        if value::kind(rec) != Kind::Set {
            return None;
        }
        let at = value::slot(rec);
        Some(self.sets.get(at)?.encoding())
    }

    /// How a hash is represented, or `None` if `key` is not a hash.
    ///
    /// The same shape as [`Keyspace::set_encoding`] and for the same reason: the
    /// record holds a slot number and the body is the thing that knows which of
    /// the two it currently is.
    pub fn hash_encoding(&mut self, key: &[u8]) -> Option<hash::Encoding> {
        self.reap(key);
        let rec = self.map.get(key)?;
        if value::kind(rec) != Kind::Hash {
            return None;
        }
        let at = value::slot(rec);
        Some(self.hashes.get(at)?.encoding())
    }

    /// How a list is represented, or `None` if `key` is not a list.
    ///
    /// The same shape as [`Keyspace::set_encoding`], and the same argument for
    /// asking the body rather than reading a copy out of the record.
    pub fn list_encoding(&mut self, key: &[u8]) -> Option<list::Encoding> {
        self.reap(key);
        let rec = self.map.get(key)?;
        if value::kind(rec) != Kind::List {
            return None;
        }
        let at = value::slot(rec);
        Some(self.lists.get(at)?.encoding())
    }

    /// How a sorted set is represented, or `None` if `key` is not one.
    pub fn zset_encoding(&mut self, key: &[u8]) -> Option<zset::Encoding> {
        self.reap(key);
        let rec = self.map.get(key)?;
        if value::kind(rec) != Kind::Zset {
            return None;
        }
        let at = value::slot(rec);
        Some(self.zsets.get(at)?.encoding())
    }

    /// `OBJECT ENCODING key`, as the word Redis puts on the wire.
    ///
    /// One place that knows every type's answer, so that adding the hash means
    /// adding an arm here and not finding the four callers that each worked it
    /// out for themselves.
    pub fn encoding_name(&mut self, key: &[u8]) -> Option<&'static str> {
        match self.kind_of(key)? {
            Kind::String => self.encoding(key).map(value::Encoding::name),
            Kind::Set => self.set_encoding(key).map(set::Encoding::name),
            Kind::Hash => self.hash_encoding(key).map(hash::Encoding::name),
            Kind::List => self.list_encoding(key).map(list::Encoding::name),
            Kind::Zset => self.zset_encoding(key).map(zset::Encoding::name),
            // The one type with one encoding, so there is nothing to ask.
            Kind::Array => Some("sliced-array"),
            // Named rather than caught, so that the next type to land is a
            // build error here and not a panic on a live server. That is not
            // hypothetical: `COPY` of a list took the shard down for exactly as
            // long as its own match had a catch all at the bottom.
            Kind::Stream => unreachable!("nothing can store a stream yet"),
        }
    }

    /// Put a deadline on `key`, or take one off. Answers whether it was there.
    ///
    /// Any type. A deadline lives in the record and changes its length, so this
    /// writes the record again rather than patching it, and for a set that is
    /// five bytes or thirteen and never the members. The body is left exactly
    /// where it is, which is why this writes through the map instead of taking
    /// the free the body path an overwrite takes.
    ///
    /// This is the raw write. [`Keyspace::expire`] and [`Keyspace::persist`] are
    /// what `EXPIRE` and its family call, and they come through here once they
    /// have worked out whether the deadline is allowed to move.
    pub fn set_expiry(&mut self, key: &[u8], at: Option<u64>) -> bool {
        self.reap(key);
        let Some(rec) = self.map.get(key) else {
            return false;
        };
        if value::expire_at(rec) == at {
            return true;
        }
        // Read what has to survive out of the record before writing over it.
        match value::kind(rec) {
            Kind::String => {
                // Through the scratch buffer rather than a fresh `Vec`, since
                // `EXPIRE` on a string is a command a cache sends as often as
                // the `SET` before it.
                let mut bytes = std::mem::take(&mut self.scratch);
                bytes.clear();
                value::read(rec).write_to(&mut bytes);
                self.store(key, &bytes, at);
                self.scratch = bytes;
            }
            // Every body type writes the same record: a tag and a slot number.
            // The body is not touched and does not need to be, which is the
            // whole point of keeping it out of the record.
            kind @ (Kind::Set | Kind::Hash | Kind::List | Kind::Zset | Kind::Array) => {
                let slot = value::slot(rec);
                let len = value::slot_record_len(at.is_some());
                self.write_rec(key, len, |out| {
                    value::write_slot_record(out, kind, slot, at);
                });
            }
            // Named rather than caught, as above.
            Kind::Stream => unreachable!("nothing can store a stream yet"),
        }
        true
    }

    /// The key's deadline, as the three way answer `TTL` and `PTTL` are built on.
    ///
    /// [`Ask::Missing`] for a key that is not there, [`Ask::NoDeadline`] for one
    /// that is and has no deadline, and the absolute millisecond otherwise. A key
    /// past its deadline is reaped on the way through, so it answers `Missing`
    /// and not the moment that has gone.
    ///
    /// Asking when a key dies is not using it, so this does not stamp the
    /// eviction clock. Redis reads the key with its no touch flag here for the
    /// same reason, and it matters more than it looks: a client polling `TTL` on
    /// a key would otherwise keep that key at the top of the working set for as
    /// long as it kept asking whether it was about to go.
    pub fn deadline_of(&mut self, key: &[u8]) -> Ask {
        let Some(addr) = self.live_rec_untouched(key) else {
            return Ask::Missing;
        };
        match value::expire_at(self.map.value_at(addr)) {
            Some(at) => Ask::At(at),
            None => Ask::NoDeadline,
        }
    }

    /// Move `key`'s deadline to `at`, if `cond` lets it.
    ///
    /// This is `EXPIRE`, `PEXPIRE`, `EXPIREAT` and `PEXPIREAT`, which differ only
    /// in the unit and the origin of the number. All four turn it into one
    /// absolute millisecond before they get here, so the condition rules live in
    /// one place and the four commands cannot drift apart.
    ///
    /// A deadline that has already passed deletes the key rather than being
    /// stored, and the answer says so. `EXPIRE` cannot report the difference
    /// because it replies 1 either way, but the caller is not always `EXPIRE`,
    /// and a delete is a different thing from a deadline.
    ///
    /// The condition is checked before the past check, which is the order Redis
    /// uses and is the one that matters: `EXPIRE key 0 XX` on a key with no
    /// deadline answers 0 and leaves the key alone, rather than deleting it.
    pub fn expire(&mut self, key: &[u8], at: u64, cond: Cond) -> Applied {
        let prev = match self.deadline_of(key) {
            Ask::Missing => return Applied::Missing,
            Ask::NoDeadline => None,
            Ask::At(at) => Some(at),
        };
        let done = ttl::decide(prev, at, cond, self.clock.now_ms());
        match done {
            Applied::Ok => {
                self.set_expiry(key, Some(at));
            }
            // The structure that answered `Deleted` for a field only holds
            // deadlines, so its caller has to remove the field. Here the caller
            // is us and the key is ours, so it goes now.
            Applied::Deleted => {
                self.drop_key(key);
            }
            Applied::Missing | Applied::NotMet => {}
        }
        done
    }

    /// Take `key`'s deadline off. Answers whether there was one to take.
    ///
    /// This is `PERSIST`, and the reply is the same 0 for a key that is not there
    /// and a key that was never going to expire, which is Redis's answer and not
    /// a shortcut here.
    pub fn persist(&mut self, key: &[u8]) -> bool {
        if !matches!(self.deadline_of(key), Ask::At(_)) {
            return false;
        }
        self.set_expiry(key, None);
        true
    }

    /// Give back whatever `key` holds outside its record, if it holds anything.
    ///
    /// Every path that deletes a key or writes over one has to come through
    /// here, because a set that loses its record without losing its slab slot is
    /// a leak that nothing ever notices: the memory is reachable, the slot is
    /// never reused, and `DBSIZE` looks right. Six delete sites and four string
    /// writers each remembering to do it themselves is five chances to forget,
    /// and one of them would be forgotten. So this is the funnel, and when the
    /// hash type lands the only place that changes is the match below.
    ///
    /// The record is left alone. This frees the body and the caller either
    /// deletes the record or writes a new one over it.
    pub(crate) fn free_body(&mut self, key: &[u8]) {
        if self.bodies == 0 {
            return;
        }
        let Some(rec) = self.map.get(key) else {
            return;
        };
        match value::kind(rec) {
            Kind::String => {}
            Kind::Set => {
                let at = value::slot(rec);
                self.sets.remove(at);
                self.bodies -= 1;
            }
            Kind::Hash => {
                let at = value::slot(rec);
                self.hashes.remove(at);
                self.bodies -= 1;
            }
            Kind::List => {
                let at = value::slot(rec);
                self.lists.remove(at);
                self.bodies -= 1;
            }
            Kind::Zset => {
                let at = value::slot(rec);
                self.zsets.remove(at);
                self.bodies -= 1;
            }
            Kind::Array => {
                let at = value::slot(rec);
                self.arrays.remove(at);
                self.bodies -= 1;
            }
            // Named rather than caught, as above.
            Kind::Stream => unreachable!("nothing can store a stream yet"),
        }
    }

    /// Delete `key` and whatever it held. Answers whether it was there.
    #[inline]
    pub(crate) fn drop_key(&mut self, key: &[u8]) -> bool {
        self.free_body(key);
        self.map.del(key)
    }

    /// Drop `key` if its deadline has passed.
    ///
    /// This is lazy expiry and it is half of the story. The other half is the
    /// active cycle in the maintenance slice, which is what stops a key nobody
    /// ever reads again from holding its memory forever (`14` section 1).
    ///
    /// Every public read calls this first, whatever type it is reading, which
    /// is why it is here and not in the file for any one type.
    #[inline]
    pub(crate) fn reap(&mut self, key: &[u8]) {
        let now = self.clock.now_ms();
        let dead = self.map.get(key).is_some_and(|r| value::is_expired(r, now));
        if dead {
            self.drop_key(key);
            self.expired += 1;
        }
    }

    /// Where `key`'s record is, having thrown the key away first if it is dead.
    ///
    /// The same fold as [`Keyspace::live_slot`] for a caller that wants the
    /// record itself rather than a slot number, which is every string command.
    /// `GET` used to be a reap, then a type check, then a read, and each of the
    /// three hashed the key and walked a bucket for the same record. It is one
    /// walk now and two arena reads, and an arena read at a known address is a
    /// load.
    ///
    /// The address dies at the next write, which is why this is `pub(crate)`
    /// and why every caller reads it and drops it inside one command.
    ///
    /// Finding a key counts as using it, so this stamps the access field on the
    /// way past under every policy that wants it stamped, which is eight of the
    /// ten. A command that has to look at a key without using it calls
    /// [`Keyspace::live_rec_untouched`] instead, and the list of those is short
    /// and is Redis's list rather than ours.
    pub(crate) fn live_rec(&mut self, key: &[u8]) -> Option<Addr> {
        let addr = self.live_rec_untouched(key)?;
        if self.policy.stamps_on_read() {
            self.stamp(addr);
        }
        Some(addr)
    }

    /// [`Keyspace::live_rec`] for a command that is asking about a key rather
    /// than using it.
    ///
    /// `TYPE`, `EXISTS`, the `TTL` family and every `OBJECT` subcommand look
    /// without touching, which is Redis's `LOOKUP_NOTOUCH` and is not an
    /// optimisation. `OBJECT IDLETIME` that counted as a use would report zero
    /// every time, and `EXISTS` in a health check loop would keep a dead key at
    /// the top of the working set forever.
    ///
    /// Untouched means the access field only. A key past its deadline is still
    /// reaped here, because a command asking whether a key exists has to be told
    /// that it does not.
    pub(crate) fn live_rec_untouched(&mut self, key: &[u8]) -> Option<Addr> {
        let now = self.clock.now_ms();
        let addr = self.map.find(key)?;
        if value::is_expired(self.map.value_at(addr), now) {
            self.drop_key(key);
            self.expired += 1;
            return None;
        }
        Some(addr)
    }

    /// The slot under `key`, having thrown the key away first if it is dead.
    ///
    /// `None` for a key that is not there or that was and is now reaped, and
    /// `WRONGTYPE` for a key holding something other than `want`.
    ///
    /// One probe of the map, where a [`Keyspace::reap`] followed by a `get`
    /// costs two. That pair is how every collection command used to start, so a
    /// pipeline of sixty four `SADD` on one key hashed and probed for that key a
    /// hundred and twenty eight times to do sixty four inserts. The reap has to
    /// read the record and the command has to read the same record, and there
    /// was never a reason for those to be two visits.
    ///
    /// It answers a number rather than the record it just read because of the
    /// borrow checker and not because a number is nicer. A method that hands
    /// back a borrow of the map on one path and takes a mutable borrow to reap
    /// on the other is the case the borrow checker still refuses without
    /// Polonius. A slot is four bytes and copies out, so the borrow ends here
    /// and the caller reaches its body through the slab.
    ///
    /// And no probe at all when the command in front of it asked for the same
    /// key and nothing has been written since, which is the [`Memo`] and is what
    /// Y13 asks for on single key `SADD`.
    pub(crate) fn live_slot(&mut self, key: &[u8], want: Kind) -> Result<Option<u32>> {
        // A memo hit skips the record, so the stamp has to happen on the way out
        // of it as well. It is the same address every time, which is what makes
        // this cheap: no probe, just the store. Getting this wrong is the trap
        // worth naming, because the key that hits the memo most often is the
        // hottest key in the database, and it is the one that would have looked
        // steadily more idle the harder it was used.
        if let Some((kind, slot, addr)) = self.memo.get(self.map.writes(), key) {
            if kind != want {
                return Err(wrong_type());
            }
            if self.policy.stamps_on_read() {
                self.stamp(addr);
            }
            return Ok(Some(slot));
        }
        // `find` and then `value_at` rather than `get`, which is the same two
        // steps, so that the address is still in hand for the stamp below. `get`
        // would mean probing a second time for a record already read.
        let now = self.clock.now_ms();
        let Some(addr) = self.map.find(key) else {
            return Ok(None);
        };
        let rec = self.map.value_at(addr);
        if value::is_expired(rec, now) {
            self.drop_key(key);
            self.expired += 1;
            return Ok(None);
        }
        if value::kind(rec) != want {
            return Err(wrong_type());
        }
        let slot = value::slot(rec);
        // A key with a deadline is not memoized. The memo is invalidated by
        // writes and a deadline passes without one, so remembering a dated key
        // would be remembering it past the moment it should have been reaped.
        // Both of these are read off the record before the stamp, which needs it
        // mutably and is the end of this borrow.
        let dated = value::expire_at(rec).is_some();
        if self.policy.stamps_on_read() {
            self.stamp(addr);
        }
        if !dated {
            self.memo.put(self.map.writes(), key, want, slot, addr);
        }
        Ok(Some(slot))
    }

    /// Where `key` is, when either of two types will do.
    ///
    /// Every input to a sorted set operation may be a sorted set or a plain set,
    /// which is Redis's rule and means the type check there is a membership test
    /// rather than an equality. The kind comes back with the slot because the
    /// caller has to know which slab the number indexes.
    pub(crate) fn live_slot_either(
        &mut self,
        key: &[u8],
        a: Kind,
        b: Kind,
    ) -> Result<Option<(Kind, u32)>> {
        if let Some((kind, slot, addr)) = self.memo.get(self.map.writes(), key) {
            if kind != a && kind != b {
                return Err(wrong_type());
            }
            if self.policy.stamps_on_read() {
                self.stamp(addr);
            }
            return Ok(Some((kind, slot)));
        }
        let now = self.clock.now_ms();
        let Some(addr) = self.map.find(key) else {
            return Ok(None);
        };
        let rec = self.map.value_at(addr);
        if value::is_expired(rec, now) {
            self.drop_key(key);
            self.expired += 1;
            return Ok(None);
        }
        let kind = value::kind(rec);
        if kind != a && kind != b {
            return Err(wrong_type());
        }
        let slot = value::slot(rec);
        let dated = value::expire_at(rec).is_some();
        if self.policy.stamps_on_read() {
            self.stamp(addr);
        }
        if !dated {
            self.memo.put(self.map.writes(), key, kind, slot, addr);
        }
        Ok(Some((kind, slot)))
    }

    /// Throw every key away. This is `FLUSHDB` on one database.
    ///
    /// The expiry counter is not reset, because Redis does not reset it either:
    /// `expired_keys` in `INFO stats` counts what this process has expired since
    /// it started, and emptying a database is not expiring anything.
    pub fn clear(&mut self) {
        self.map.clear();
        self.sets.clear();
        self.hashes.clear();
        self.lists.clear();
        self.zsets.clear();
        self.arrays.clear();
        self.bodies = 0;
    }

    /// Keys reclaimed by running into them after their deadline.
    ///
    /// Redis calls this `expired_keys` in `INFO stats` and counts both lazy and
    /// active expiry into it. Only lazy expiry exists so far, so only lazy
    /// expiry is counted, and the active cycle in the maintenance slice will add
    /// to the same number when it lands (`14` section 1).
    #[inline]
    pub const fn expired_keys(&self) -> u64 {
        self.expired
    }

    /// Keys thrown away to make room.
    ///
    /// Redis calls this `evicted_keys` in `INFO stats`. It stays at zero under
    /// `noeviction`, which is the whole point of that policy, and a monitoring
    /// dashboard that sees it move on a server configured that way is looking at
    /// a bug rather than at load.
    #[inline]
    pub const fn evicted_keys(&self) -> u64 {
        self.evicted
    }

    /// How many keys a round of eviction sampling looks at.
    #[inline]
    pub const fn samples(&self) -> usize {
        self.samples
    }

    /// Set how many keys a round of eviction sampling looks at.
    ///
    /// Zero is not refused here, because the caller doing the refusing is
    /// `CONFIG SET` and it has a message to produce. A zero that reaches here
    /// samples one bucket and takes the best of it, because the loop runs its
    /// body before it checks, which is a better answer than dividing by nothing.
    #[inline]
    pub const fn set_samples(&mut self, samples: usize) {
        self.samples = samples;
    }

    /// Throw away one key, chosen by the policy. Answers whether one went.
    ///
    /// This is one step and not a loop on purpose. The caller is the thing that
    /// knows how much room it needs back, and a loop in here would either take
    /// too much or have to be told the same number twice. It also means the
    /// caller can put a bound on how long it spends evicting before it answers
    /// the client, which matters because the client is waiting on a write that
    /// this is making room for.
    ///
    /// It answers false without doing anything under `noeviction`, and also when
    /// a `volatile` policy is set on a database where nothing has a deadline.
    /// Those are the same answer to the caller and they mean the same thing: this
    /// server cannot give memory back and is about to have to refuse a write.
    pub fn evict_one(&mut self) -> bool {
        let Some(addr) = self.victim() else {
            return false;
        };
        // The key has to outlive the borrow that found it, because deleting is a
        // write and the address came out of a read. One copy into the scratch
        // buffer rather than a `Vec` per eviction, for the reason written on
        // [`Keyspace::scratch`]: this runs in a loop when it runs at all.
        let mut buf = core::mem::take(&mut self.scratch);
        buf.clear();
        buf.extend_from_slice(self.map.entry_at(addr).0);
        let gone = self.drop_key(&buf);
        self.scratch = buf;
        if gone {
            self.evicted += 1;
        }
        gone
    }

    /// Where the key this policy would throw away lives, if there is one.
    ///
    /// The sampling loop. It draws buckets until it has looked at `samples` keys
    /// the policy would consider, scores each one, and hands back the best. See
    /// [`evict`] for what the score means and [`yo_index::RawMap::sample`] for
    /// why a bucket is the unit.
    ///
    /// The round cap is the part that is not obvious. A database with a hundred
    /// keys in a directory sized for a million is mostly empty buckets, and a
    /// `volatile` policy on a database where nothing has a deadline has no
    /// eligible keys at all however many buckets it looks in. Without the cap the
    /// second case is an infinite loop, and it is not a rare configuration, it is
    /// the classic eviction surprise. With it, the worst case is a fixed number
    /// of cache misses and a false, which is exactly what the caller needs to
    /// hear.
    ///
    /// A key past its deadline is skipped rather than taken. It is dead memory
    /// and evicting it would look like a win, but it would be counted as an
    /// eviction when it is an expiry, and those two numbers are watched
    /// separately for a reason. Lazy expiry takes it the next time anything asks
    /// for it, and the active cycle takes it before that.
    fn victim(&mut self) -> Option<Addr> {
        if matches!(self.policy, Policy::NoEviction) || self.map.is_empty() {
            return None;
        }
        let now = self.clock.now_ms();
        let (policy, lfu, want) = (self.policy, self.lfu, self.samples);
        let mut best = evict::Best::EMPTY;
        let mut seen = 0usize;
        for _ in 0..ROUNDS {
            let r = self.rng.next_u64();
            self.map.sample(r, |_key, rec, addr| {
                if !value::is_expired(rec, now) && evict::eligible(rec, policy) {
                    seen += 1;
                    best.offer(addr, evict::score(rec, policy, now, lfu));
                }
                seen < want
            });
            if seen >= want {
                break;
            }
        }
        (!best.is_empty()).then_some(best.addr)
    }

    /// Bytes held by the index, the arena and every body hanging off them.
    ///
    /// Asks every collection, so this is O(the number of collections) and is for
    /// the places that want the number exactly and are asked for it rarely:
    /// `INFO memory`, `MEMORY USAGE` and the tests.
    /// [`Keyspace::settled_memory_bytes`] is the one a memory limit uses.
    #[inline]
    pub fn memory_bytes(&self) -> usize {
        self.slab_bytes()
            + self.sets.value_bytes()
            + self.hashes.value_bytes()
            + self.lists.value_bytes()
            + self.zsets.value_bytes()
            + self.arrays.value_bytes()
    }

    /// The same number, asked only of the collections that could have moved.
    ///
    /// See [`Slab::track_bytes`] for how that is known. With tracking on this
    /// costs what the batch touched instead of what the database holds, which is
    /// what lets a server with a `maxmemory` ask once a batch. With tracking off
    /// it is [`Keyspace::memory_bytes`] and the two cannot disagree, because
    /// they are the same sum over the same values either way.
    #[inline]
    pub fn settled_memory_bytes(&mut self) -> usize {
        self.slab_bytes()
            + self.sets.settled_bytes()
            + self.hashes.settled_bytes()
            + self.lists.settled_bytes()
            + self.zsets.settled_bytes()
            + self.arrays.settled_bytes()
    }

    /// Start or stop keeping the running total in every slab.
    ///
    /// One call for all five, because a limit is a property of the server and
    /// not of a type, and a database tracking its sets but not its hashes would
    /// answer a number that is neither of the two things it could mean.
    pub fn track_memory(&mut self, on: bool) {
        self.sets.track_bytes(on);
        self.hashes.track_bytes(on);
        self.lists.track_bytes(on);
        self.zsets.track_bytes(on);
        self.arrays.track_bytes(on);
    }

    /// The index, the arena and the slot arrays, none of which need asking twice.
    #[inline]
    fn slab_bytes(&self) -> usize {
        self.map.memory_bytes()
            + self.sets.slot_bytes()
            + self.hashes.slot_bytes()
            + self.lists.slot_bytes()
            + self.zsets.slot_bytes()
            + self.arrays.slot_bytes()
    }

    /// Give back one segment's worth of space if one has gone mostly dead.
    ///
    /// Overwriting a key does not reuse its bytes, it writes the new record at
    /// the bump pointer and counts the old one as dead, so a workload that sets
    /// the same keys over and over holds far more than it is storing until
    /// something compacts. This is that something, and it does at most one
    /// segment per call so that the loop can afford to ask every turn.
    #[inline]
    pub fn compact_step(&mut self) -> Option<usize> {
        self.map.compact_step()
    }

    /// The same, for a store that is over a memory limit and has to give pages
    /// back rather than wait for a segment to be worth collecting.
    ///
    /// See [`RawMap::compact_hard`] for why the choice of segment changes and
    /// why it only changes under pressure.
    #[inline]
    pub fn compact_hard(&mut self) -> Option<usize> {
        self.map.compact_hard()
    }

    /// Ask the cache for the bucket this key will land in.
    ///
    /// The first of the loop's two walks (`04` section 3) calls this.
    #[inline]
    pub fn prefetch(&self, hash: u64) {
        self.map.prefetch(hash);
    }

    /// The hash this database files `key` under.
    #[inline]
    #[must_use]
    pub fn hash_of(key: &[u8]) -> u64 {
        RawMap::hash_of(key)
    }
}

/// What Redis says when a command is sent at a key holding another type.
///
/// The text is Redis's, word for word, because it goes on the wire verbatim and
/// clients match on it. The `WRONGTYPE` at the front is not part of the message:
/// the protocol layer puts it there from the [`Code`], which is what lets an
/// embedded caller match on a value instead of on a string (P5).
pub fn wrong_type() -> Error {
    Error::new(
        Code::WrongType,
        "Operation against a key holding the wrong kind of value",
    )
}

impl Default for Keyspace {
    fn default() -> Keyspace {
        Keyspace::new()
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    fn db() -> Keyspace {
        Keyspace::with_clock(Clock::fixed(1_000))
    }

    #[test]
    fn type_answers_string_for_a_string_and_nothing_for_a_missing_key() {
        let mut d = db();
        d.set_plain(b"k", b"v").expect("room");
        assert_eq!(d.kind_of(b"k"), Some(Kind::String));
        assert_eq!(d.kind_of(b"nope"), None);
    }

    #[test]
    fn type_does_not_report_a_key_whose_deadline_has_gone() {
        let mut d = db();
        d.psetex(b"k", 100, b"v").expect("room");
        assert_eq!(d.kind_of(b"k"), Some(Kind::String));

        d.clock_mut().advance(100);
        assert_eq!(
            d.kind_of(b"k"),
            None,
            "the deadline was 1100 and it is 1100"
        );
        assert_eq!(d.len(), 0, "and asking reaped it rather than leaving it");
        assert_eq!(d.expired_keys(), 1);
    }

    /// The default policy evicts nothing and still keeps the clock, which is
    /// Redis's behaviour and is the configuration nearly every server runs.
    #[test]
    fn the_clock_runs_under_the_default_policy() {
        let mut d = db();
        assert_eq!(d.policy(), Policy::NoEviction);
        d.set_plain(b"k", b"v").expect("room");
        assert_eq!(d.idle_secs(b"k"), Some(0));

        d.clock_mut().advance(60_000);
        assert_eq!(d.idle_secs(b"k"), Some(60), "a minute of nobody asking");

        d.get(b"k").expect("a string").expect("still there");
        assert_eq!(d.idle_secs(b"k"), Some(0), "and reading it is using it");
    }

    /// The commands that ask about a key rather than use it. Getting this wrong
    /// makes `OBJECT IDLETIME` answer zero every time it is called, because
    /// calling it would be the most recent use.
    #[test]
    fn asking_about_a_key_is_not_using_it() {
        let mut d = db();
        d.set_plain(b"k", b"v").expect("room");
        d.clock_mut().advance(30_000);

        assert!(d.exists(b"k"));
        assert_eq!(d.kind_of(b"k"), Some(Kind::String));
        assert_eq!(d.encoding_name(b"k"), Some("embstr"));
        assert_eq!(d.deadline_of(b"k"), Ask::NoDeadline);
        assert_eq!(d.expire_at(b"k"), None);
        assert_eq!(d.idle_secs(b"k"), Some(30));

        assert_eq!(
            d.idle_secs(b"k"),
            Some(30),
            "and asking twice is still not using it"
        );
    }

    /// Least recently modified is the one policy where a read must leave the
    /// clock where it is, because the clock is the only thing it measures.
    #[test]
    fn a_read_moves_the_clock_under_lru_and_leaves_it_under_lrm() {
        for (policy, idle_after_read) in [(Policy::AllKeysLru, 0), (Policy::AllKeysLrm, 45)] {
            let mut d = db();
            d.set_policy(policy);
            d.set_plain(b"k", b"v").expect("room");
            d.clock_mut().advance(45_000);

            d.get(b"k").expect("a string").expect("still there");
            assert_eq!(
                d.idle_secs(b"k"),
                Some(idle_after_read),
                "{}",
                policy.name()
            );

            // Both of them move it on a write, which is the whole of what LRM
            // is measuring and is a side effect of the resolve under LRU.
            d.set_plain(b"k", b"w").expect("room");
            assert_eq!(
                d.idle_secs(b"k"),
                Some(0),
                "{} after a write",
                policy.name()
            );
        }
    }

    /// The trap the memo sets. A hit skips the record entirely, so a stamp that
    /// only happened on a miss would leave the hottest key in the database
    /// looking steadily more idle the harder it was used.
    #[test]
    fn the_hot_key_path_still_stamps() {
        let mut d = db();
        d.set_policy(Policy::AllKeysLru);
        d.sadd(b"s", [&b"a"[..]].into_iter()).expect("room");

        // Warm the memo, then run the key hard with nothing written in between,
        // which is the case the memo exists for.
        d.scard(b"s").expect("a set");
        d.clock_mut().advance(120_000);
        for _ in 0..64 {
            d.scard(b"s").expect("a set");
        }
        assert_eq!(d.idle_secs(b"s"), Some(0), "the memo swallowed the stamp");
    }

    /// Under LFU the same bits are a counter, and it climbs with use rather than
    /// resetting to now.
    #[test]
    fn the_counter_climbs_under_an_lfu_policy() {
        let mut d = db();
        d.set_policy(Policy::AllKeysLfu);
        d.seed(7);
        d.set_plain(b"k", b"v").expect("room");
        let start = d.freq(b"k").expect("there");

        for _ in 0..200 {
            d.get(b"k").expect("a string").expect("still there");
        }
        let hot = d.freq(b"k").expect("there");
        assert!(hot > start, "{hot} did not climb from {start}");

        // And a key nobody reads decays rather than holding its place forever.
        d.set_plain(b"cold", b"v").expect("room");
        d.clock_mut().advance(60_000 * 10);
        assert!(d.freq(b"cold").expect("there") < start);
    }

    /// Nothing goes under `noeviction`, which is the only promise that policy
    /// makes and the reason it is the default.
    #[test]
    fn noeviction_evicts_nothing() {
        let mut d = db();
        for i in 0..200u32 {
            d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
        }
        assert!(!d.evict_one());
        assert_eq!(d.len(), 200);
        assert_eq!(d.evicted_keys(), 0);
    }

    /// A volatile policy on a database where nothing has a deadline is the
    /// classic surprise: it looks configured and it cannot free a byte.
    #[test]
    fn a_volatile_policy_with_no_deadlines_anywhere_cannot_evict() {
        let mut d = db();
        d.set_policy(Policy::VolatileLru);
        for i in 0..200u32 {
            d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
        }
        assert!(!d.evict_one(), "it found a key it had no business taking");
        assert_eq!(d.len(), 200);

        // Give one key a deadline and it becomes the only thing that can go,
        // however many rounds of sampling that takes.
        let deadline = d.clock().now_ms() + 100_000;
        d.set_expiry(b"k7", Some(deadline));
        assert!(d.evict_one());
        assert!(!d.exists(b"k7"));
        assert_eq!(d.evicted_keys(), 1);
    }

    /// The direction of the score, which is the thing worth pinning. A test that
    /// only checked something was evicted would pass just as happily on a cache
    /// that keeps the cold keys and throws away the hot ones.
    #[test]
    fn the_stale_key_goes_before_the_fresh_one() {
        let mut d = db();
        d.set_policy(Policy::AllKeysLru);
        // Two keys is a small enough database that a bucket holds both of them
        // and the pick is between them rather than between whatever turned up.
        d.set_plain(b"cold", b"v").expect("room");
        d.clock_mut().advance(600_000);
        d.set_plain(b"hot", b"v").expect("room");

        assert!(d.evict_one());
        assert!(!d.exists(b"cold"), "it kept the stale one");
        assert!(d.exists(b"hot"), "it took the fresh one");
    }

    /// Under `volatile-ttl` the ordering is by deadline and not by use, so the
    /// key about to expire anyway is the one that goes.
    #[test]
    fn the_soonest_deadline_goes_first() {
        let mut d = db();
        d.set_policy(Policy::VolatileTtl);
        let now = d.clock().now_ms();
        d.set_plain(b"soon", b"v").expect("room");
        d.set_plain(b"later", b"v").expect("room");
        d.set_expiry(b"soon", Some(now + 10_000));
        d.set_expiry(b"later", Some(now + 900_000));

        assert!(d.evict_one());
        assert!(!d.exists(b"soon"));
        assert!(d.exists(b"later"));
    }

    /// Under LFU the key nobody reads goes, even though it was written more
    /// recently than the one that survives. That is the difference between the
    /// two families and it is invisible to a test written against the clock.
    #[test]
    fn the_least_used_key_goes_under_lfu() {
        let mut d = db();
        d.set_policy(Policy::AllKeysLfu);
        d.seed(11);
        d.set_plain(b"popular", b"v").expect("room");
        for _ in 0..300 {
            d.get(b"popular").expect("a string").expect("still there");
        }
        // Written after the reads above, so under any clock policy this would be
        // the freshest key in the database and the last thing to go.
        d.set_plain(b"ignored", b"v").expect("room");

        assert!(d.evict_one());
        assert!(!d.exists(b"ignored"));
        assert!(d.exists(b"popular"));
    }

    /// Sampling has to keep working when almost every bucket it looks in is
    /// empty, which is what a database looks like after most of it is deleted.
    #[test]
    fn a_nearly_empty_database_still_gives_up_a_key() {
        let mut d = db();
        d.set_policy(Policy::AllKeysRandom);
        for i in 0..4000u32 {
            d.set_plain(format!("k{i}").as_bytes(), b"v").expect("room");
        }
        for i in 0..3999u32 {
            d.drop_key(format!("k{i}").as_bytes());
        }
        assert_eq!(d.len(), 1);

        // One key in a directory sized for four thousand. It may take more than
        // one round to land on it, and it may take more than one call, but the
        // rounds are bounded and so is this loop.
        let mut went = false;
        for _ in 0..500 {
            if d.evict_one() {
                went = true;
                break;
            }
        }
        assert!(went, "sampling never found the one key that was left");
        assert_eq!(d.len(), 0);
        assert!(!d.evict_one(), "and an empty database has nothing to give");
    }

    /// Eviction and expiry are counted apart, so a key that was already dead
    /// when sampling found it is not billed as an eviction.
    #[test]
    fn a_dead_key_is_not_evicted() {
        let mut d = db();
        d.set_policy(Policy::AllKeysLru);
        let now = d.clock().now_ms();
        d.set_plain(b"k", b"v").expect("room");
        d.set_expiry(b"k", Some(now + 1000));
        d.clock_mut().advance(5000);

        assert!(!d.evict_one(), "it evicted a key that was already dead");
        assert_eq!(d.evicted_keys(), 0);
    }
}