950 lines
34 KiB
Rust
950 lines
34 KiB
Rust
//! Where the time goes when a narrowed content filter scrubs the index.
|
|
//!
|
|
//! Narrowing `content_extensions` sets `IndexWork { reconcile_content: true }`
|
|
//! (`config::diff_actions`). `scope::advance` then reads every stored row
|
|
//! under every root, re-decides each one, and for each row whose verdict
|
|
//! changed clears its content and flips `files.content_state`. This probe
|
|
//! attributes that time to a phase and prices the shapes that could replace it.
|
|
//!
|
|
//! ```text
|
|
//! cargo build -p quicksearch-core --example contentprobe --release
|
|
//! ./target/release/examples/contentprobe
|
|
//! ```
|
|
//!
|
|
//! `QSB_FILES`, `QSB_DIRS`, `QSB_DIR_DEPTH` and `QSB_PENDING_EVERY` size the
|
|
//! corpus; `QSB_ARMS=plain|keyed|both` picks the key states (see
|
|
//! `common::arms`).
|
|
//!
|
|
//! # The sibling probe
|
|
//!
|
|
//! `pruneprobe` measures the other half of the same function: an added ignore
|
|
//! pattern, which deletes rows outright. The two differ in how much of what
|
|
//! they touch holds a posting. A narrowed content filter targets rows that do
|
|
//! (95% of the rewritten rows on this corpus, against 14% of the doomed rows
|
|
//! on pruneprobe's), so this path is the slower one and gains the most from
|
|
//! the `deletemerge` lever below.
|
|
//!
|
|
//! The stages are cumulative, so consecutive rows subtract to a phase cost:
|
|
//!
|
|
//! ```text
|
|
//! page read the rows, decide nothing
|
|
//! +decide ...and re-run content_extractable per row
|
|
//! +state ...and flip content_state, per row
|
|
//! +clear(all) ...and clear FTS/text/failure per row
|
|
//! +clear(done) ...clearing only rows that can hold one
|
|
//! +chunked ...issuing the clears as chunked IN (...) lists
|
|
//! +chunk state ...and the state flip chunked too
|
|
//! ```
|
|
//!
|
|
//! # What it found
|
|
//!
|
|
//! The FTS5 tombstone dominates the clears: 786 ms of a 975 ms pass, against
|
|
//! 13 ms for `documents_text` and 1 ms for `failed_files`. The table splits
|
|
//! the three because one `clear` column would suggest fixing the wrong thing.
|
|
//!
|
|
//! Statement shape barely matters, and is easy to misread. Chunking the clears
|
|
//! takes `fts` from 786 ms to 19 ms and puts 848 ms into `commit`, for 6% off
|
|
//! `total`: SQLite spills the work to whichever statement is executing when
|
|
//! the page cache fills, so compare only `total` across shapes. `+chunk state`
|
|
//! demonstrates it, moving the same ~850 ms onto an `UPDATE` that the query
|
|
//! plans show is a plain rowid seek.
|
|
//!
|
|
//! `deletemerge` is the lever, worth 6-7x; read the knob sweep.
|
|
//! `file_handling::fts_begin_tombstone_burst` owns the result and ships it,
|
|
//! so `live` lands far under every stage.
|
|
|
|
mod common;
|
|
|
|
use std::path::Path;
|
|
use std::time::{Duration, Instant};
|
|
|
|
use rusqlite::Connection;
|
|
|
|
use quicksearch_core::config::Config;
|
|
use quicksearch_core::db::repo;
|
|
use quicksearch_core::extract::Registry;
|
|
use quicksearch_core::file_handling::{content_extractable, RootRange, FTS_DELETEMERGE};
|
|
use quicksearch_core::scope::{self, WorkCursor};
|
|
use quicksearch_core::testutil::{self, Arm, SeedSpec};
|
|
|
|
use common::{
|
|
chunked, clone_arm, count, env_usize, merge_to_quiescence, open, placeholders, print_merge,
|
|
set_cache, Commit, Io, CHUNK, MERGE_PAGES, PAGE, SHIPPED_MIB,
|
|
};
|
|
|
|
/// The root every `seed_index` row hangs under. It need not exist on disk:
|
|
/// the content re-decision stats no stored path, reading the extension off
|
|
/// the string and the MIME off the row.
|
|
const ROOT: &str = "/seed";
|
|
|
|
/// The corpus's extensions and the MIME each carries.
|
|
///
|
|
/// Eleven, coprime with `dirs` (2000 by default). `seed_index` assigns
|
|
/// extensions by `i % len` and directories by `i % dirs`, so a shared factor
|
|
/// would confine each extension to some of the directories: the rewritten rows
|
|
/// would arrive in a few dense runs, the easy case for page locality, and
|
|
/// flatter every figure. `main` asserts it.
|
|
const EXT_MIX: &[(&str, &str)] = &[
|
|
("txt", "text/plain"),
|
|
("jpg", "image/jpeg"),
|
|
("png", "image/png"),
|
|
("md", "text/markdown"),
|
|
("zip", "application/zip"),
|
|
("mp4", "video/mp4"),
|
|
("pdf", "application/pdf"),
|
|
("gif", "image/gif"),
|
|
("bin", "application/octet-stream"),
|
|
("mov", "video/quicktime"),
|
|
("ttf", "font/ttf"),
|
|
];
|
|
|
|
/// The [`EXT_MIX`] entries an extractor claims, and so the rows seeded with a
|
|
/// posting: 3 of 11.
|
|
///
|
|
/// [`NARROWED`] and the fraction sweep are written in terms of these three, so
|
|
/// `main` checks them against `Registry::default_set()`: an extractor gaining
|
|
/// or losing a MIME fails there instead of reshaping every figure.
|
|
const EXTRACTABLE: [&str; 3] = ["txt", "md", "pdf"];
|
|
|
|
/// The filter the stage table narrows to, from an unfiltered index. Keeping
|
|
/// `txt` and dropping `md` and `pdf` rewrites 2/11 of all rows and tombstones
|
|
/// 2/3 of the postings; the fraction sweep covers a filter that takes every
|
|
/// posting.
|
|
const NARROWED: &[&str] = &["txt"];
|
|
|
|
/// The `files.name` test for the rows [`NARROWED`] withdraws content from,
|
|
/// for the census.
|
|
const NARROWED_AWAY: &str = "(name LIKE '%.md' OR name LIKE '%.pdf')";
|
|
|
|
/// Greatest common divisor, for the stride assertion in `main`.
|
|
fn gcd(a: usize, b: usize) -> usize {
|
|
if b == 0 {
|
|
a
|
|
} else {
|
|
gcd(b, a % b)
|
|
}
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Stages
|
|
// ---------------------------------------------------------------------------
|
|
|
|
/// Cumulative slices of the re-decision. Each does everything the one before
|
|
/// it does.
|
|
#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
|
|
enum Stage {
|
|
Page,
|
|
Decide,
|
|
State,
|
|
ClearAll,
|
|
ClearDone,
|
|
Chunked,
|
|
ChunkedState,
|
|
}
|
|
|
|
impl Stage {
|
|
const ALL: [Stage; 7] = [
|
|
Stage::Page,
|
|
Stage::Decide,
|
|
Stage::State,
|
|
Stage::ClearAll,
|
|
Stage::ClearDone,
|
|
Stage::Chunked,
|
|
Stage::ChunkedState,
|
|
];
|
|
|
|
fn label(self) -> &'static str {
|
|
match self {
|
|
Stage::Page => "page",
|
|
Stage::Decide => "+decide",
|
|
Stage::State => "+state",
|
|
Stage::ClearAll => "+clear(all)",
|
|
Stage::ClearDone => "+clear(done)",
|
|
Stage::Chunked => "+chunked",
|
|
Stage::ChunkedState => "+chunk state",
|
|
}
|
|
}
|
|
|
|
fn tag(self) -> &'static str {
|
|
match self {
|
|
Stage::Page => "page",
|
|
Stage::Decide => "decide",
|
|
Stage::State => "state",
|
|
Stage::ClearAll => "clear-all",
|
|
Stage::ClearDone => "clear-done",
|
|
Stage::Chunked => "chunked",
|
|
Stage::ChunkedState => "chunked-state",
|
|
}
|
|
}
|
|
|
|
/// Whether the re-decision runs at all.
|
|
fn decides(self) -> bool {
|
|
self >= Stage::Decide
|
|
}
|
|
|
|
/// Whether `files.content_state` is written.
|
|
fn writes_state(self) -> bool {
|
|
self >= Stage::State
|
|
}
|
|
|
|
/// Whether the row's content is cleared alongside the state flip.
|
|
fn clears(self) -> bool {
|
|
self >= Stage::ClearAll
|
|
}
|
|
|
|
/// Whether the clear is narrowed to rows whose stored state says they can
|
|
/// hold something to clear.
|
|
fn clear_filtered(self) -> bool {
|
|
self >= Stage::ClearDone
|
|
}
|
|
|
|
/// Whether the clears go out as chunked `IN (...)` lists instead of one
|
|
/// bound statement per row.
|
|
fn chunked_clears(self) -> bool {
|
|
self >= Stage::Chunked
|
|
}
|
|
|
|
/// Whether the `content_state` flip is chunked too. A separate step
|
|
/// because it does not behave like the clears; see the query plans `main`
|
|
/// prints.
|
|
fn chunked_state(self) -> bool {
|
|
self >= Stage::ChunkedState
|
|
}
|
|
}
|
|
|
|
/// What one stage cost, split by phase.
|
|
#[derive(Default)]
|
|
struct Timing {
|
|
page: Duration,
|
|
decide: Duration,
|
|
state: Duration,
|
|
/// The three clears, timed apart: one phase, three different pieces of
|
|
/// work. An FTS5 contentless delete writes a tombstone into every segment
|
|
/// covering the row's origin, a `documents_text` delete frees a compressed
|
|
/// body's overflow pages, and a `failed_files` delete seeks into a table
|
|
/// that is nearly always empty.
|
|
fts: Duration,
|
|
text: Duration,
|
|
failed: Duration,
|
|
commit: Duration,
|
|
total: Duration,
|
|
examined: usize,
|
|
/// Rows whose content was withdrawn.
|
|
to_na: usize,
|
|
/// Rows put back in the pending queue.
|
|
to_pending: usize,
|
|
/// Bytes the block layer served this pass. The arm was just copied, so
|
|
/// near zero means the working set was in memory and every duration is
|
|
/// CPU and page-cache work.
|
|
read_bytes: u64,
|
|
misses: i64,
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// The measured loop
|
|
// ---------------------------------------------------------------------------
|
|
|
|
/// The ids one page decided to write, split by what each one needs.
|
|
#[derive(Default)]
|
|
struct Pending {
|
|
/// Rows moving to `STATE_NA`.
|
|
to_na: Vec<i64>,
|
|
/// Rows moving to `STATE_PENDING`.
|
|
to_pending: Vec<i64>,
|
|
/// Of the two above, those whose *stored* state was `STATE_DONE` and so
|
|
/// can hold a posting and a stored body.
|
|
had_posting: Vec<i64>,
|
|
/// ...and those whose stored state was `STATE_FAILED`, the only ones that
|
|
/// can hold a `failed_files` row.
|
|
had_failure: Vec<i64>,
|
|
/// Every id above, in the order the per-row loop visits them.
|
|
all: Vec<i64>,
|
|
}
|
|
|
|
impl Pending {
|
|
fn clear(&mut self) {
|
|
self.to_na.clear();
|
|
self.to_pending.clear();
|
|
self.had_posting.clear();
|
|
self.had_failure.clear();
|
|
self.all.clear();
|
|
}
|
|
|
|
fn is_empty(&self) -> bool {
|
|
self.all.is_empty()
|
|
}
|
|
|
|
/// Sort every list ascending, which for these ids is rowid order; see
|
|
/// [`Knobs::sort_ids`].
|
|
fn sort(&mut self) {
|
|
for list in [
|
|
&mut self.to_na,
|
|
&mut self.to_pending,
|
|
&mut self.had_posting,
|
|
&mut self.had_failure,
|
|
&mut self.all,
|
|
] {
|
|
list.sort_unstable();
|
|
}
|
|
}
|
|
}
|
|
|
|
/// Levers over FTS5's tombstone machinery, applied for the length of the pass.
|
|
/// Neither changes what the index ends up holding: the final `'merge'`
|
|
/// reclaims the tombstones either way.
|
|
#[derive(Clone, Copy, Default)]
|
|
struct Knobs {
|
|
/// Set FTS5's `deletemerge` to 0 for the pass and restore it after. At its
|
|
/// default of 10, each level that passes 10% tombstones is rewritten
|
|
/// inline with the scan; at 0 they accumulate for the final merge. See
|
|
/// `file_handling::fts_begin_tombstone_burst`.
|
|
no_deletemerge: bool,
|
|
/// Sort each page's ids ascending before deleting.
|
|
///
|
|
/// The scan serves rows in `(parent, name)` order, uncorrelated with
|
|
/// rowid, so each `%_docsize` lookup in `fts5StorageContentlessDelete`
|
|
/// lands somewhere new. Sorting restores locality within a page, not
|
|
/// across pages. Measured and rejected; see
|
|
/// `file_handling::fts_begin_tombstone_burst`.
|
|
sort_ids: bool,
|
|
}
|
|
|
|
impl Knobs {
|
|
fn label(self) -> String {
|
|
match (self.no_deletemerge, self.sort_ids) {
|
|
(false, false) => "neither".into(),
|
|
(true, false) => "deletemerge off".into(),
|
|
(false, true) => "rowid-sorted".into(),
|
|
(true, true) => "both".into(),
|
|
}
|
|
}
|
|
|
|
fn apply(self, conn: &Connection) {
|
|
if self.no_deletemerge {
|
|
set_deletemerge(conn, 0);
|
|
}
|
|
}
|
|
|
|
/// Undo [`Knobs::apply`]. FTS5 persists `deletemerge` in `%_config`, so a
|
|
/// leaked 0 would change what every later reading of the file means,
|
|
/// including the merge that follows.
|
|
fn restore(self, conn: &Connection) {
|
|
if self.no_deletemerge {
|
|
set_deletemerge(conn, i64::from(FTS_DELETEMERGE));
|
|
}
|
|
}
|
|
}
|
|
|
|
fn set_deletemerge(conn: &Connection, percent: i64) {
|
|
conn.execute(
|
|
"INSERT INTO searchabletext(searchabletext, rank) VALUES('deletemerge', ?1)",
|
|
[percent],
|
|
)
|
|
.expect("set deletemerge");
|
|
}
|
|
|
|
/// Page the index as `scope::advance` does, doing `stage`'s share of the work
|
|
/// and timing each phase apart.
|
|
///
|
|
/// The writes are spelled out here rather than routed through `repo`, so the
|
|
/// per-row and chunked variants compare in one run and the stages stay fixed
|
|
/// when `repo` changes.
|
|
fn run_stage(
|
|
conn: &Connection,
|
|
config: &Config,
|
|
registry: &Registry,
|
|
stage: Stage,
|
|
commit: Commit,
|
|
) -> Timing {
|
|
run_stage_with(conn, config, registry, stage, commit, Knobs::default())
|
|
}
|
|
|
|
fn run_stage_with(
|
|
conn: &Connection,
|
|
config: &Config,
|
|
registry: &Registry,
|
|
stage: Stage,
|
|
commit: Commit,
|
|
knobs: Knobs,
|
|
) -> Timing {
|
|
let range = RootRange::for_root(ROOT);
|
|
let max_size = config.processing.maximum_text_file_size;
|
|
|
|
let mut t = Timing::default();
|
|
knobs.apply(conn);
|
|
let io_before = Io::read();
|
|
let (_, misses_before) = testutil::cache_stats(conn);
|
|
let started = Instant::now();
|
|
|
|
let mut after = (range.lo.clone(), String::new());
|
|
let mut pending = Pending::default();
|
|
// `unchecked_transaction` borrows shared, so one transaction can end and
|
|
// the next begin inside the loop, as in `testutil::seed_index`.
|
|
let mut tx = conn.unchecked_transaction().expect("begin");
|
|
let mut pages_open = 0usize;
|
|
let mut tx_since = Instant::now();
|
|
loop {
|
|
let at = Instant::now();
|
|
let rows = repo::rows_in_range_page(&tx, &after.0, &after.1, &range.hi, PAGE)
|
|
.expect("read a page");
|
|
t.page += at.elapsed();
|
|
let Some(last) = rows.last() else { break };
|
|
after = (last.parent.clone(), last.name.clone());
|
|
pages_open += 1;
|
|
t.examined += rows.len();
|
|
|
|
if !stage.decides() {
|
|
continue;
|
|
}
|
|
|
|
// `scope::apply_page`'s decision.
|
|
let at = Instant::now();
|
|
pending.clear();
|
|
for row in &rows {
|
|
let path = Path::new(&row.path);
|
|
let wants = row.size <= max_size
|
|
&& content_extractable(path, row.mime.as_deref(), config, registry);
|
|
if !wants && row.content_state != repo::STATE_NA {
|
|
pending.to_na.push(row.id);
|
|
} else if wants && row.content_state == repo::STATE_NA {
|
|
pending.to_pending.push(row.id);
|
|
} else {
|
|
continue;
|
|
}
|
|
pending.all.push(row.id);
|
|
if row.content_state == repo::STATE_DONE {
|
|
pending.had_posting.push(row.id);
|
|
} else if row.content_state == repo::STATE_FAILED {
|
|
pending.had_failure.push(row.id);
|
|
}
|
|
}
|
|
if knobs.sort_ids {
|
|
pending.sort();
|
|
}
|
|
t.decide += at.elapsed();
|
|
t.to_na += pending.to_na.len();
|
|
t.to_pending += pending.to_pending.len();
|
|
|
|
if stage.writes_state() && !pending.is_empty() {
|
|
write_page(&tx, stage, &pending, &mut t);
|
|
}
|
|
|
|
if commit.due(pages_open, tx_since) {
|
|
let at = Instant::now();
|
|
tx.commit().expect("commit");
|
|
t.commit += at.elapsed();
|
|
tx = conn.unchecked_transaction().expect("begin");
|
|
pages_open = 0;
|
|
tx_since = Instant::now();
|
|
}
|
|
}
|
|
let at = Instant::now();
|
|
tx.commit().expect("commit");
|
|
t.commit += at.elapsed();
|
|
|
|
t.total = started.elapsed();
|
|
t.read_bytes = Io::read().since(&io_before).read_bytes;
|
|
let (_, misses_after) = testutil::cache_stats(conn);
|
|
t.misses = misses_after - misses_before;
|
|
knobs.restore(conn);
|
|
t
|
|
}
|
|
|
|
/// Apply one page's decisions, in `stage`'s shape, timing the clear and the
|
|
/// state flip apart.
|
|
fn write_page(tx: &Connection, stage: Stage, pending: &Pending, t: &mut Timing) {
|
|
// The ids the clear is offered: those whose stored state can hold
|
|
// something, or every flipped row.
|
|
let (fts_ids, failed_ids): (&[i64], &[i64]) = if stage.clear_filtered() {
|
|
(&pending.had_posting, &pending.had_failure)
|
|
} else {
|
|
(&pending.all, &pending.all)
|
|
};
|
|
|
|
// The per-row clears are `repo::remove_content_for_id` followed by a
|
|
// `failed_files` delete, one bound statement at a time. They prepare through the connection's cache, as
|
|
// `repo::exec` does; preparing afresh per row would price a mistake the
|
|
// product does not make.
|
|
if stage.clears() {
|
|
// Each target runs to completion, timed on its own, before the next
|
|
// starts: interleaving them as `remove_content_for_id` does would
|
|
// charge whichever statement was executing when the page cache
|
|
// spilled.
|
|
for (in_list, per_row, ids, spent) in [
|
|
(
|
|
"DELETE FROM searchabletext WHERE rowid IN",
|
|
"DELETE FROM searchabletext WHERE rowid = ?1",
|
|
fts_ids,
|
|
&mut t.fts,
|
|
),
|
|
(
|
|
"DELETE FROM documents_text WHERE file_id IN",
|
|
"DELETE FROM documents_text WHERE file_id = ?1",
|
|
fts_ids,
|
|
&mut t.text,
|
|
),
|
|
(
|
|
"DELETE FROM failed_files WHERE file_id IN",
|
|
"DELETE FROM failed_files WHERE file_id = ?1",
|
|
failed_ids,
|
|
&mut t.failed,
|
|
),
|
|
] {
|
|
let at = Instant::now();
|
|
if stage.chunked_clears() {
|
|
chunked(tx, in_list, ids, true);
|
|
} else {
|
|
for id in ids {
|
|
one(tx, per_row, *id);
|
|
}
|
|
}
|
|
*spent += at.elapsed();
|
|
}
|
|
}
|
|
|
|
let at = Instant::now();
|
|
for (state, ids) in [
|
|
(repo::STATE_NA, &pending.to_na),
|
|
(repo::STATE_PENDING, &pending.to_pending),
|
|
] {
|
|
if stage.chunked_state() {
|
|
chunked_state(tx, state, ids);
|
|
} else {
|
|
for id in ids {
|
|
tx.prepare_cached("UPDATE files SET content_state = ?1 WHERE id = ?2")
|
|
.and_then(|mut s| s.execute(rusqlite::params![state, id]))
|
|
.expect("flip content_state");
|
|
}
|
|
}
|
|
}
|
|
t.state += at.elapsed();
|
|
}
|
|
|
|
fn one(tx: &Connection, sql: &str, id: i64) {
|
|
tx.prepare_cached(sql)
|
|
.and_then(|mut s| s.execute([id]))
|
|
.expect("per-row write");
|
|
}
|
|
|
|
fn chunked_state(tx: &Connection, state: i64, ids: &[i64]) {
|
|
for chunk in ids.chunks(CHUNK) {
|
|
let sql = format!(
|
|
"UPDATE files SET content_state = ?1 WHERE id IN ({})",
|
|
placeholders(chunk.len())
|
|
);
|
|
let params = std::iter::once(&state).chain(chunk.iter());
|
|
tx.prepare_cached(&sql)
|
|
.and_then(|mut s| s.execute(rusqlite::params_from_iter(params)))
|
|
.expect("chunked state flip");
|
|
}
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Corpus
|
|
// ---------------------------------------------------------------------------
|
|
|
|
fn spec() -> SeedSpec {
|
|
SeedSpec {
|
|
files: env_usize("QSB_FILES", 200_000),
|
|
dirs: env_usize("QSB_DIRS", 2_000),
|
|
// Must be 1. With a mix, the row's MIME decides extractability and
|
|
// `seed_index` intersects that with this stride, so any other value
|
|
// leaves rows an extractor claims at `STATE_NA`: the first reconcile
|
|
// would repair the seed instead of applying the edit.
|
|
content_every: 1,
|
|
// One in twenty extractable rows never got its content, the residue
|
|
// of an interrupted run. The unfiltered clears still delete FTS,
|
|
// stored text and a failure record for these rows, though none can
|
|
// exist; `+clear(done)` prices that.
|
|
pending_every: env_usize("QSB_PENDING_EVERY", 20),
|
|
ext_mix: EXT_MIX,
|
|
// Row width, and so pages scanned; see `SeedSpec::dir_depth`.
|
|
dir_depth: env_usize("QSB_DIR_DEPTH", 6).max(2),
|
|
// A real run commits in slices, each flushing FTS5's hash to its own
|
|
// segment; one transaction would leave a single segment and
|
|
// understate every clear cost.
|
|
commit_every: 500,
|
|
..SeedSpec::default()
|
|
}
|
|
}
|
|
|
|
fn config_with(extensions: &[&str]) -> Config {
|
|
let mut config = Config::default();
|
|
config.paths.indexing_paths = vec![ROOT.to_string()];
|
|
config.indexing.content_extensions = extensions.iter().map(|e| e.to_string()).collect();
|
|
config.processing.batch_size = PAGE as usize;
|
|
config
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
|
|
fn main() {
|
|
let spec = spec();
|
|
let registry = Registry::default_set();
|
|
|
|
// Check the corpus's premise.
|
|
let claimed: Vec<&str> = EXT_MIX
|
|
.iter()
|
|
.filter(|(_, mime)| registry.supports(mime))
|
|
.map(|(ext, _)| *ext)
|
|
.collect();
|
|
assert_eq!(
|
|
claimed, EXTRACTABLE,
|
|
"the registry no longer claims exactly the extensions this probe is \
|
|
written around; NARROWED and the fraction sweep name them by hand"
|
|
);
|
|
assert_eq!(
|
|
gcd(EXT_MIX.len(), spec.dirs),
|
|
1,
|
|
"the extension stride ({}) shares a factor with the directory stride \
|
|
({}), so each extension reaches only some of the directories — the \
|
|
rewritten rows would cluster instead of spreading. See EXT_MIX",
|
|
EXT_MIX.len(),
|
|
spec.dirs
|
|
);
|
|
|
|
println!(
|
|
"corpus: {} files across {} dirs, {} of {} extensions extractable ({}), \
|
|
1 in {} of those left pending; narrowing to {:?}",
|
|
spec.files,
|
|
spec.dirs,
|
|
EXTRACTABLE.len(),
|
|
EXT_MIX.len(),
|
|
EXTRACTABLE.join("/"),
|
|
spec.pending_every,
|
|
NARROWED
|
|
);
|
|
|
|
for keyed in common::arms() {
|
|
let label = if keyed { "keyed" } else { "plain" };
|
|
let master = Arm::seed(label, &format!("content-{}", label), keyed, &spec);
|
|
println!(
|
|
"\n=== {} === seeded in {:.1}s, {}",
|
|
label,
|
|
master.seeded_in.as_secs_f64(),
|
|
common::mib(master.size_bytes())
|
|
);
|
|
|
|
// Census: what the narrowing touches, read off the index.
|
|
{
|
|
let conn = open(&master);
|
|
let total = count(&conn, "SELECT COUNT(*) FROM files");
|
|
let done = count(&conn, "SELECT COUNT(*) FROM files WHERE content_state = 1");
|
|
let pending = count(&conn, "SELECT COUNT(*) FROM files WHERE content_state = 0");
|
|
let na = count(&conn, "SELECT COUNT(*) FROM files WHERE content_state = 3");
|
|
let fts = count(&conn, "SELECT COUNT(*) FROM searchabletext");
|
|
assert_eq!(fts, done, "a posting exists exactly for content_state = 1");
|
|
assert!(
|
|
pending > 0,
|
|
"no pending residue — QSB_PENDING_EVERY is 0, and +clear(done) \
|
|
has nothing to skip"
|
|
);
|
|
// The rows the narrowing rewrites: everything not already NA whose
|
|
// extension leaves the filter.
|
|
let rewritten = count(
|
|
&conn,
|
|
&format!(
|
|
"SELECT COUNT(*) FROM files WHERE content_state != 3 AND {}",
|
|
NARROWED_AWAY
|
|
),
|
|
);
|
|
let with_posting = count(
|
|
&conn,
|
|
&format!(
|
|
"SELECT COUNT(*) FROM files WHERE content_state = 1 AND {}",
|
|
NARROWED_AWAY
|
|
),
|
|
);
|
|
assert!(
|
|
rewritten > 0,
|
|
"the narrowing rewrites nothing — seed_index's naming moved"
|
|
);
|
|
assert!(
|
|
with_posting < rewritten,
|
|
"every rewritten row holds a posting, so +clear(done) has \
|
|
nothing to skip and the stage cannot answer anything"
|
|
);
|
|
println!(
|
|
" {} rows: {} done, {} pending, {} n/a. {} rewritten by the \
|
|
narrowing ({:.0}%), of which {} hold a posting ({:.0}%) — \
|
|
the other {} hold nothing to clear",
|
|
total,
|
|
done,
|
|
pending,
|
|
na,
|
|
rewritten,
|
|
100.0 * rewritten as f64 / total as f64,
|
|
with_posting,
|
|
100.0 * with_posting as f64 / rewritten as f64,
|
|
rewritten - with_posting,
|
|
);
|
|
// The query plans, which a timing cannot show: whether each shape
|
|
// seeks or scans.
|
|
for (what, sql) in [
|
|
(
|
|
"delete, per row ",
|
|
"DELETE FROM searchabletext WHERE rowid = 1".to_string(),
|
|
),
|
|
(
|
|
"delete, chunked ",
|
|
"DELETE FROM searchabletext WHERE rowid IN (1,2,3)".to_string(),
|
|
),
|
|
(
|
|
"state, per row ",
|
|
"UPDATE files SET content_state = 3 WHERE id = 1".to_string(),
|
|
),
|
|
(
|
|
"state, chunked ",
|
|
"UPDATE files SET content_state = 3 WHERE id IN (1,2,3)".to_string(),
|
|
),
|
|
] {
|
|
let plan: Vec<String> = conn
|
|
.prepare(&format!("EXPLAIN QUERY PLAN {}", sql))
|
|
.and_then(|mut s| {
|
|
s.query_map([], |r| r.get::<_, String>(3))?
|
|
.collect::<rusqlite::Result<Vec<_>>>()
|
|
})
|
|
.unwrap_or_else(|e| vec![format!("unavailable: {}", e)]);
|
|
println!(" plan {}: {}", what, plan.join(" | "));
|
|
}
|
|
common::print_fts_census(&conn);
|
|
}
|
|
|
|
let config = config_with(NARROWED);
|
|
let header = || {
|
|
println!(
|
|
"\n {:<16} {:>6} {:>6} {:>6} {:>7} {:>6} {:>6} {:>7} {:>7} {:>8} {:>7}",
|
|
"stage",
|
|
"page",
|
|
"decide",
|
|
"state",
|
|
"fts",
|
|
"text",
|
|
"failed",
|
|
"commit",
|
|
"total",
|
|
"misses",
|
|
"written"
|
|
);
|
|
};
|
|
let row = |name: &str, t: &Timing| {
|
|
let ms = |d: Duration| d.as_secs_f64() * 1000.0;
|
|
println!(
|
|
" {:<16} {:>6.0} {:>6.0} {:>6.0} {:>7.0} {:>6.0} {:>6.0} {:>7.0} {:>7.0} {:>8} {:>7}",
|
|
name,
|
|
ms(t.page),
|
|
ms(t.decide),
|
|
ms(t.state),
|
|
ms(t.fts),
|
|
ms(t.text),
|
|
ms(t.failed),
|
|
ms(t.commit),
|
|
ms(t.total),
|
|
t.misses,
|
|
t.to_na + t.to_pending,
|
|
);
|
|
};
|
|
|
|
header();
|
|
for stage in Stage::ALL {
|
|
let arm = clone_arm(&master, &format!("content-{}-{}", label, stage.tag()));
|
|
let conn = open(&arm);
|
|
let t = run_stage(&conn, &config, ®istry, stage, Commit::PerPage);
|
|
row(stage.label(), &t);
|
|
// Consolidation, priced apart: merge until `%_data` stops changing,
|
|
// as `fts_end_tombstone_burst` ends a burst.
|
|
if stage.clears() {
|
|
print_merge(&conn, 16, "");
|
|
}
|
|
drop(conn);
|
|
arm.discard();
|
|
}
|
|
|
|
// Is the cost per commit or per row? The first line of each block
|
|
// commits per page; if the cost collapses as pages fold together,
|
|
// commit cadence is the lever, not the write shape. Both shapes are
|
|
// swept because the two fixes are independent and either could
|
|
// subsume the other.
|
|
for stage in [Stage::ClearAll, Stage::Chunked] {
|
|
println!("\n transaction size, at {}:", stage.label());
|
|
header();
|
|
for pages in [1usize, 8, 64, 512] {
|
|
let arm = clone_arm(
|
|
&master,
|
|
&format!("content-{}-{}-tx{}", label, stage.tag(), pages),
|
|
);
|
|
let conn = open(&arm);
|
|
let t = run_stage(&conn, &config, ®istry, stage, Commit::Pages(pages));
|
|
row(&format!("{} page/tx", pages), &t);
|
|
drop(conn);
|
|
arm.discard();
|
|
}
|
|
let arm = clone_arm(&master, &format!("content-{}-{}-slice", label, stage.tag()));
|
|
let conn = open(&arm);
|
|
let t = run_stage(
|
|
&conn,
|
|
&config,
|
|
®istry,
|
|
stage,
|
|
Commit::Slice(scope::SLICE),
|
|
);
|
|
row(&format!("{:?} slice", scope::SLICE), &t);
|
|
drop(conn);
|
|
arm.discard();
|
|
}
|
|
|
|
// FTS5's own machinery, where the time is once statement shape stops
|
|
// mattering. Run at the chunked shape, so what remains is FTS5's cost
|
|
// and not per-row overhead.
|
|
println!("\n FTS5 knobs, at +chunked (+ merge to quiescence):");
|
|
header();
|
|
for knobs in [
|
|
Knobs::default(),
|
|
Knobs {
|
|
no_deletemerge: true,
|
|
..Knobs::default()
|
|
},
|
|
Knobs {
|
|
sort_ids: true,
|
|
..Knobs::default()
|
|
},
|
|
Knobs {
|
|
no_deletemerge: true,
|
|
sort_ids: true,
|
|
},
|
|
] {
|
|
let arm = clone_arm(
|
|
&master,
|
|
&format!("content-{}-k{}", label, knobs.label().len()),
|
|
);
|
|
let conn = open(&arm);
|
|
let t = run_stage_with(
|
|
&conn,
|
|
&config,
|
|
®istry,
|
|
Stage::Chunked,
|
|
Commit::Slice(scope::SLICE),
|
|
knobs,
|
|
);
|
|
row(&knobs.label(), &t);
|
|
print_merge(&conn, 16, "");
|
|
drop(conn);
|
|
arm.discard();
|
|
}
|
|
|
|
// The page cache. `PRAGMAS_INCREMENTAL` gives the reconcile a flat
|
|
// 4 MiB for a pass across the whole index. The scan reads
|
|
// `idx_files_parent` in order but fetches a table row per entry, and
|
|
// the writes scatter into `%_docsize`, `documents_text` and FTS5's
|
|
// tombstone pages in orders uncorrelated with `(parent, name)`. Read
|
|
// the `misses` column.
|
|
println!(
|
|
"\n page cache, at +clear(all) [{} = shipped]:",
|
|
SHIPPED_MIB
|
|
);
|
|
header();
|
|
for mib in [SHIPPED_MIB, 16, 32, 64, 128, 256] {
|
|
for (name, commit) in [
|
|
("per page", Commit::PerPage),
|
|
("per slice", Commit::Slice(scope::SLICE)),
|
|
] {
|
|
let arm = clone_arm(
|
|
&master,
|
|
&format!("content-{}-c{}-{}", label, mib, name.len()),
|
|
);
|
|
let conn = open(&arm);
|
|
set_cache(&conn, mib);
|
|
let t = run_stage(&conn, &config, ®istry, Stage::ClearAll, commit);
|
|
row(&format!("{} MiB, {}", mib, name), &t);
|
|
drop(conn);
|
|
arm.discard();
|
|
}
|
|
}
|
|
|
|
// The withdrawn fraction, swept. Past some fraction FTS5 stops
|
|
// tombstoning and starts rewriting segments, which costs what building
|
|
// the index did. `baseline` is `+clear(all)` committing per page;
|
|
// `fixed` is `+chunked` committing per slice.
|
|
println!("\n withdrawn fraction (baseline -> fixed, both + merge to quiescence):");
|
|
println!(
|
|
" {:<13} {:>8} {:>9} {:>9} {:>9} {:>11} {:>11}",
|
|
"keeping", "written", "baseline", "fixed", "merge", "%_data", "postings"
|
|
);
|
|
for keep in [&["txt", "md"][..], &["txt"][..], &["rtf"][..]] {
|
|
let config = config_with(keep);
|
|
let tag = keep.join("+");
|
|
let mut timings = Vec::new();
|
|
for (what, stage, commit) in [
|
|
("baseline", Stage::ClearAll, Commit::PerPage),
|
|
("fixed", Stage::Chunked, Commit::Slice(scope::SLICE)),
|
|
] {
|
|
let arm = clone_arm(&master, &format!("content-{}-{}-{}", label, tag, what));
|
|
let conn = open(&arm);
|
|
let before = common::fts_data_rows(&conn);
|
|
let t = run_stage(&conn, &config, ®istry, stage, commit);
|
|
let (merge, _, _, after) = merge_to_quiescence(&conn, MERGE_PAGES);
|
|
let left = count(&conn, "SELECT COUNT(*) FROM searchabletext");
|
|
timings.push((t, merge, before, after, left));
|
|
drop(conn);
|
|
arm.discard();
|
|
}
|
|
let ms = |d: Duration| d.as_secs_f64() * 1000.0;
|
|
println!(
|
|
" {:<13} {:>8} {:>8.0} {:>8.0} {:>8.0} {:>11} {:>11}",
|
|
tag,
|
|
timings[0].0.to_na + timings[0].0.to_pending,
|
|
ms(timings[0].0.total),
|
|
ms(timings[1].0.total),
|
|
ms(timings[1].1),
|
|
format!("{}->{}", timings[1].2, timings[1].3),
|
|
timings[1].4,
|
|
);
|
|
}
|
|
|
|
// The real `scope::advance`, driven to completion as the coordinator
|
|
// drives it. It lands under every stage; see the module doc.
|
|
{
|
|
let arm = clone_arm(&master, &format!("content-{}-live", label));
|
|
let mut conn = open(&arm);
|
|
let unfiltered = config_with(&[]);
|
|
let actions = quicksearch_core::config::diff_actions(&unfiltered, &config);
|
|
assert!(
|
|
actions.work.reconcile_content && !actions.work.reindex,
|
|
"the narrowing should reconcile content and ask for no walk, got {:?}",
|
|
actions.work
|
|
);
|
|
let mut cursor = WorkCursor::new(actions.work, &config).expect("plan");
|
|
let run = std::sync::atomic::AtomicBool::new(false);
|
|
let started = Instant::now();
|
|
while !cursor.done() {
|
|
scope::advance(
|
|
&mut conn,
|
|
&config,
|
|
®istry,
|
|
&mut cursor,
|
|
Instant::now() + scope::SLICE,
|
|
&run,
|
|
)
|
|
.expect("advance");
|
|
}
|
|
println!(
|
|
" {:<16} {:>6} {:>6} {:>6} {:>7} {:>6} {:>6} {:>7} {:>7.0} {:>8} {:>7} <- scope::advance",
|
|
"live",
|
|
"",
|
|
"",
|
|
"",
|
|
"",
|
|
"",
|
|
"",
|
|
"",
|
|
started.elapsed().as_secs_f64() * 1000.0,
|
|
"",
|
|
cursor.recontented,
|
|
);
|
|
drop(conn);
|
|
arm.discard();
|
|
}
|
|
|
|
master.discard();
|
|
}
|
|
}
|