kevy_embedded/

ops_table.rs

//! Embedded `TABLE.*` capability — the declaration
//! object, the compile call and the verify sweep. The grammar, the
//! validation and [`kevy_index::TableSpec::compile`] all live in
//! `kevy-index`: ONE implementation shared with the server, so the two
//! engines cannot compile a table differently (the IDX.CREATE parity
//! lesson; the dispatch oracle byte-compares the wire faces anyway).

use std::sync::{Mutex, RwLock};

use kevy_index::{AdviseLog, IndexVerify, TableCatalog, TableEnsure, TableSpec, TableVerify};

use crate::store::{Store, lock_write};
use crate::{KevyError, KevyResult};

/// Store-level table state (declarations only — a table's runtime
/// footprint is its compiled indexes in the index registry).
#[derive(Debug, Default)]
pub(crate) struct TableReg {
    pub(crate) catalog: RwLock<TableCatalog>,
    /// The refusal log (the auto-declaration loop's observation
    /// face) — fed by the typed query API's refusals, rendered by
    /// [`Store::idx_advise`], cleared on every catalog mutation.
    pub(crate) advise: Mutex<AdviseLog>,
}

/// Rows the per-shard column spot check samples (mirrors the server).
const SPOTCHECK_ROWS: usize = 64;

/// One `TABLE.VERIFY` result: per compiled index its name + six
/// counters (entries, bytes, coerce_failures, duplicates, drift,
/// checked), plus the `(rows, type_mismatches)` spot-check pair.
/// The legacy verify shape: six unnamed counters per index and two for
/// the spot check. Kept for semver; superseded by [`TableVerify`],
/// whose fields carry the names and time semantics these arrays never
/// could (dogfood F10).
#[deprecated(since = "4.1.0", note = "use `table_verify_report`, whose fields are named")]
pub type TableVerifyReport = (Vec<(Vec<u8>, [u64; 6])>, [u64; 2]);

impl Store {
    /// `TABLE.DECLARE` equivalent: admit the declaration, then build
    /// every compiled index synchronously. Atomic against catalog
    /// errors: names are dry-run against a catalog clone first, so a
    /// collision installs nothing.
    pub fn table_declare(&self, spec: TableSpec) -> KevyResult<()> {
        self.catalog_change(|| self.declare_table(spec))
    }

    fn declare_table(&self, spec: TableSpec) -> KevyResult<()> {
        // Tiering floor refusal — the same precheck IDX.CREATE runs,
        // moved ahead of the catalog mutation so a refused declare
        // installs nothing.
        #[cfg(all(feature = "tier", not(target_arch = "wasm32")))]
        crate::ops_index_sync::tier_floor_check(&self.shards)?;
        // TableSpec::compile validates for itself — a bad spec is a named
        // refusal here, never a panic downstream (dogfood F9).
        let compiled = spec.compile().map_err(|e| KevyError::InvalidInput(e.to_string()))?;
        {
            let g = self.tables.catalog.read().unwrap_or_else(std::sync::PoisonError::into_inner);
            if g.get(&spec.name).is_some() {
                return Err(KevyError::InvalidInput("table already exists".into()));
            }
        }
        // Dry-run the compiled specs against a clone of the index
        // catalog — the server admits into a clone and installs once;
        // this is the same all-or-nothing shape.
        {
            let g = self.indexes.catalog.read().unwrap_or_else(std::sync::PoisonError::into_inner);
            let mut probe = g.1.clone();
            for ispec in &compiled {
                probe.create(ispec.clone()).map_err(|e| KevyError::InvalidInput(e.to_string()))?;
            }
        }
        {
            let mut g =
                self.tables.catalog.write().unwrap_or_else(std::sync::PoisonError::into_inner);
            g.create(spec).map_err(|e| KevyError::InvalidInput(e.to_string()))?;
        }
        for ispec in compiled {
            self.register_spec(ispec)?;
        }
        self.advise_clear();
        Ok(())
    }

    /// `TABLE.ENSURE` equivalent — the boot verb.
    ///
    /// The dogfood report's F8.2: declaring at boot is the steady
    /// state, and plain `table_declare` punishes it — a re-declare is
    /// an error, so the obvious code (declare, log the error at debug)
    /// keeps running against whatever shape the *first* boot declared,
    /// forever, silently. `ensure` gives the boot path its true verb:
    /// an identical spec is a no-op success, a changed spec is a named
    /// refusal carrying what changed — never a silent rebuild, because
    /// a rebuild is a full backfill and must be asked for by name
    /// ([`Self::table_replace`]).
    pub fn table_ensure(&self, spec: TableSpec) -> KevyResult<TableEnsure> {
        let existing = {
            let g = self.tables.catalog.read().unwrap_or_else(std::sync::PoisonError::into_inner);
            g.get(&spec.name).cloned()
        };
        match existing {
            None => {
                self.table_declare(spec)?;
                Ok(TableEnsure::Created)
            }
            Some(cur) if cur.sans_auto() == spec => Ok(TableEnsure::Unchanged),
            Some(cur) => {
                Err(KevyError::InvalidInput(kevy_index::spec_diff(&cur.sans_auto(), &spec)))
            }
        }
    }

    /// `TABLE.REPLACE` equivalent — drop and redeclare, rebuilding
    /// every compiled index from the rows. Explicitly named because it
    /// is a full backfill: the cost is asked for, not implied. The new
    /// spec is compiled (and therefore validated) *before* the old
    /// table is dropped, so a bad replacement leaves the old one
    /// standing.
    pub fn table_replace(&self, spec: TableSpec) -> KevyResult<()> {
        spec.compile().map_err(|e| KevyError::InvalidInput(e.to_string()))?;
        self.catalog_change(|| {
            self.drop_table(&spec.name);
            self.declare_table(spec)
        })
    }

    /// `TABLE.DROP` equivalent — drops the table AND its compiled
    /// indexes; `false` if absent. Refused on a replica and after
    /// [`Store::shutdown`], like every write.
    pub fn table_drop(&self, name: &[u8]) -> KevyResult<bool> {
        self.catalog_change(|| Ok(self.drop_table(name)))
    }

    fn drop_table(&self, name: &[u8]) -> bool {
        let compiled: Vec<Vec<u8>> = {
            let g = self.tables.catalog.read().unwrap_or_else(std::sync::PoisonError::into_inner);
            g.get(name)
                .map(|s| {
                    s.compile()
                        .map(|c| c.into_iter().map(|i| i.name().to_vec()).collect())
                        .unwrap_or_default() // catalog entries were admitted validated
                })
                .unwrap_or_default()
        };
        let hit = {
            let mut g =
                self.tables.catalog.write().unwrap_or_else(std::sync::PoisonError::into_inner);
            g.drop_table(name)
        };
        if hit {
            for iname in &compiled {
                self.drop_index(iname);
            }
            self.advise_clear();
        }
        hit
    }

    /// Declared tables, declaration order.
    pub fn table_list(&self) -> Vec<TableSpec> {
        let g = self.tables.catalog.read().unwrap_or_else(std::sync::PoisonError::into_inner);
        g.iter().cloned().collect()
    }

    /// `TABLE.VERIFY` equivalent: each compiled index's full drift
    /// recheck (the IDX.VERIFY discipline — composites re-derive their
    /// byte encoding) plus a bounded column-type spot check, both
    /// through the no-promote peek.
    #[deprecated(since = "4.1.0", note = "use `table_verify_report`, whose fields are named")]
    #[allow(deprecated)]
    pub fn table_verify(&self, name: &[u8]) -> KevyResult<TableVerifyReport> {
        let r = self.table_verify_report(name)?;
        Ok((
            r.per_index
                .into_iter()
                .map(|i| {
                    (
                        i.name,
                        [
                            i.entries,
                            i.approx_bytes,
                            i.coerce_failures,
                            i.duplicates,
                            i.drift,
                            i.checked,
                        ],
                    )
                })
                .collect(),
            [r.spot_rows, r.spot_type_mismatches],
        ))
    }

    /// `TABLE.VERIFY`, with every counter named and its time semantics
    /// documented on the field (see [`IndexVerify`]).
    pub fn table_verify_report(&self, name: &[u8]) -> KevyResult<TableVerify> {
        let spec = {
            let g = self.tables.catalog.read().unwrap_or_else(std::sync::PoisonError::into_inner);
            g.get(name).cloned()
        }
        .ok_or_else(|| KevyError::NotFound("no such table".into()))?;
        let compiled = spec.compile().map_err(|e| KevyError::InvalidInput(e.to_string()))?;
        let mut per_index: Vec<(Vec<u8>, [u64; 10])> =
            compiled.iter().map(|i| (i.name().to_vec(), [0u64; 10])).collect();
        let mut spot = [0u64; 2];
        for shard in self.shards.iter() {
            let mut g = lock_write(shard);
            let inner = &mut *g;
            crate::ops_index_sync::sync_segs(&self.indexes, &mut inner.idx_segs, &mut inner.store);
            for (slot, ispec) in per_index.iter_mut().zip(&compiled) {
                shard_index_counts(inner, ispec, &mut slot.1);
            }
            let (r, m) = shard_spot_check(&mut inner.store, &spec);
            spot[0] += r;
            spot[1] += m;
        }
        let mut report = TableVerify::default();
        report.per_index = per_index
            .into_iter()
            .map(|(name, s)| {
                let mut v = IndexVerify::new(name);
                (v.entries, v.approx_bytes, v.coerce_failures) = (s[0], s[1], s[2]);
                (v.duplicates, v.drift, v.checked) = (s[3], s[4], s[5]);
                (v.excluded, v.absent, v.rows, v.missing) = (s[6], s[7], s[8], s[9]);
                v
            })
            .collect();
        (report.spot_rows, report.spot_type_mismatches) = (spot[0], spot[1]);
        Ok(report)
    }
}

/// One shard's contribution to one compiled index's verify counters —
/// both directions.
///
/// index→row (`drift`): every held entry re-derives from its row.
/// row→index: every prefix row classifies against the spec,
/// with the cause kept — `absent` (NULL, by design), `coerce_failures`
/// (present-but-wrong-type, fresh: the 4.0 lifetime counter of this
/// name also swallowed absences), `excluded` (composite oversize, the
/// silent two-row gap of dogfood F8/F9, now a named number), and
/// `missing` (derives a value, has no entry — the class a drift walk
/// structurally cannot see, and the one behind F13/F14's forgotten
/// writers).
///
/// Layout: [entries, bytes, coerce_fresh, duplicates, drift, checked,
///          excluded, absent, rows, missing].
/// The row→index half of the walk, under the no-promote peek: every
/// prefix row classified by cause (the server's `classify_prefix_rows`
/// is the byte-parity twin).
/// Returns `[coerce_fresh, excluded, absent, rows, missing]`.
/// `window` carries the cold side's own count — the twin of the
/// server's `cmd_table_verify::classify_prefix_rows`, including its
/// reason: a row that slid is absent from the hot entries by design,
/// but only if it actually reached a segment, so the rows below the
/// boundary are reconciled against the live cold entries instead of
/// being exempted for sitting there.
fn classify_prefix_rows(
    s: &mut kevy_store::Store,
    spec: &kevy_index::IndexSpec,
    row_keys: &[Vec<u8>],
    indexed: &std::collections::HashSet<&[u8]>,
    window: Option<kevy_index::WindowAudit>,
) -> [u64; 5] {
    let names = spec.scalar_read_names();
    let w = spec.primary_width();
    let mut f = [0u64; 5];
    let mut below = 0u64;
    for key in row_keys {
        f[3] += 1;
        let cls = match s.peek_hash_fields(key, &names[..w]) {
            Ok(Some(vals)) => spec.classify_scalar(&vals),
            // A non-hash or vanished row has no columns: NULL row.
            _ => kevy_index::RowDerivation::Absent,
        };
        match cls {
            kevy_index::RowDerivation::Indexed(_) => {
                let slid = window.is_some_and(|wa| {
                    let v = match s.peek_hash_fields(key, &names[..w]) {
                        Ok(Some(vals)) => spec.derive_scalar(&vals),
                        _ => None,
                    };
                    v.and_then(|v| v.window_value(wa.shape)).is_some_and(|wv| wv < wa.boundary)
                });
                if !indexed.contains(key.as_slice()) {
                    if slid {
                        below += 1;
                    } else {
                        f[4] += 1;
                    }
                }
            }
            kevy_index::RowDerivation::CoerceFailed => f[0] += 1,
            kevy_index::RowDerivation::Oversize => f[1] += 1,
            kevy_index::RowDerivation::Absent => f[2] += 1,
        }
    }
    // Every row that slid should have an entry waiting for it; the
    // shortfall is the rows that fell between the tree and the segment.
    f[4] += below.saturating_sub(window.map_or(0, |w| w.cold_live));
    f
}

/// The window boundary a `missing` count must not fault rows against:
/// a row below it slid into a cold segment on purpose. `None` when the
/// path is unwindowed or has not slid yet (and always on wasm, where
/// nothing slides).
fn hot_floor_of(
    inner: &crate::store_inner::Inner,
    name: &[u8],
    ty: kevy_index::ValType,
) -> Option<kevy_index::WindowAudit> {
    #[cfg(not(target_arch = "wasm32"))]
    {
        inner
            .idx_segs
            .windows
            .iter()
            .find(|(n, _)| n.as_slice() == name)
            .and_then(|(_, w)| w.audit(ty))
    }
    #[cfg(target_arch = "wasm32")]
    {
        let _ = (inner, name, ty);
        None
    }
}

fn shard_index_counts(
    inner: &mut crate::store_inner::Inner,
    ispec: &kevy_index::IndexSpec,
    sums: &mut [u64; 10],
) {
    let Some((spec, seg)) = inner.idx_segs.segs.iter().find(|(s, _)| s.name() == ispec.name())
    else {
        return;
    };
    let stats = seg.stats();
    let mut entries: Vec<(Vec<u8>, kevy_index::IndexValue)> = Vec::new();
    seg.each_entry(|k, v| entries.push((k.to_vec(), v.clone())));
    let indexed: std::collections::HashSet<&[u8]> =
        entries.iter().map(|(k, _)| k.as_slice()).collect();
    let spec = spec.clone();
    let mut pat = spec.prefix().to_vec();
    pat.push(b'*');
    let row_keys = inner.store.collect_keys(Some(&pat), None);
    let window = hot_floor_of(inner, spec.name(), spec.ty());
    let store = &mut inner.store;
    let (drift, fresh) = store.peek_scope(|s| {
        let names = spec.scalar_read_names();
        let w = spec.primary_width();
        let mut drift = 0u64;
        for (key, held) in &entries {
            let actual = match s.peek_hash_fields(key, &names[..w]) {
                Ok(Some(vals)) => spec.derive_scalar(&vals),
                _ => None,
            };
            if actual.as_ref() != Some(held) {
                drift += 1;
            }
        }
        (drift, classify_prefix_rows(s, &spec, &row_keys, &indexed, window))
    });
    sums[0] += stats.entries;
    sums[1] += stats.approx_bytes;
    sums[2] += fresh[0];
    sums[3] += stats.duplicates;
    sums[4] += drift;
    sums[5] += entries.len() as u64;
    sums[6] += fresh[1];
    sums[7] += fresh[2];
    sums[8] += fresh[3];
    sums[9] += fresh[4];
}

/// Sample up to [`SPOTCHECK_ROWS`] rows on one shard: every PRESENT
/// declared-typed column must coerce (absent = NULL — Law 3; a
/// non-hash row counts as a row with no columns).
fn shard_spot_check(store: &mut kevy_store::Store, spec: &TableSpec) -> (u64, u64) {
    let mut pat = spec.prefix.clone();
    pat.push(b'*');
    let keys = store.collect_keys(Some(&pat), Some(SPOTCHECK_ROWS));
    let names: Vec<&[u8]> = spec.columns.iter().map(|(n, _)| n.as_slice()).collect();
    store.peek_scope(|s| {
        let (mut rows, mut mismatches) = (0u64, 0u64);
        for key in &keys {
            rows += 1;
            let Ok(Some(vals)) = s.peek_hash_fields(key, &names) else { continue };
            for ((_, ty), val) in spec.columns.iter().zip(&vals) {
                if let Some(raw) = val
                    && kevy_index::IndexValue::coerce(*ty, raw).is_none()
                {
                    mismatches += 1;
                }
            }
        }
        (rows, mismatches)
    })
}