Skip to content

SQLite rivals, rmq and River uncapped, READMEs in benchmark-repo shape, five datasets - #4

Merged
fahimfaisaal merged 9 commits into
mainfrom
feat/sqlite-rivals
Oct 25, 2026
Merged

fahimfaisaal merged 9 commits into
mainfrom
feat/sqlite-rivals

Conversation

@fahimfaisaal

@fahimfaisaal fahimfaisaal commented Oct 24, 2026 •

Copy link
Copy Markdown
Member

What

  • sqlite suite: VarMQ + sqliteq now races goqite v0.4.0, backlite v0.8.0 and River v0.49.0 (riversqlite driver). Rows are BenchmarkSQLite/<scenario>/ckpt<n>/<subject>, 32 rows, run 2026-10-25. Every subject opens SQLite with sqliteq's connection settings and the row's wal_autocheckpoint. Forced settings: River's per-queue FetchCooldown at its 1 ms minimum (the 100 ms default caps 16 workers at 160 jobs/s whatever the database does), River's logger discarded, backlite's ReleaseAfter (no default) at 10 min. walMB is the total of the run's WAL files, so one file per queue is measured like one file for all.
  • redis suite: rmq polls every 1 ms instead of the example's 100 ms, which sat as a cap on its numbers (5.9k–10.3k jobs/s at C1, 102 ms pickup latency; now 27k–33k and 1.5 ms). Full rerun 2026-10-25, 150 rows.
  • site: five datasets instead of six: Worker Pool (24/8/4 CPUs), Message Queue Redis (id redis; C1, Enqueue and Latency groups with VarMQ and the four rivals, C4 groups with VarMQDistributed and the four), Message Queue SQLite (id sqlite). Old ?id=redis-persistent|redis-distributed|sqlite-persistent links stop working.
  • docs: root, pool, redis and sqlite READMEs cut to results, subjects with versions, matrix, running; every number generated from the committed txt files. The redis per-library source notes move to redis/NOTES.md, the deploy detail to worker/README.md. LICENSE gains the goptics line next to the pond-benchmarks one.

Verified

  • cd sqlite && go test ./...: every subject through TestSubjects.
  • task site:json && jq -c '[.id,.name,(.data|length)]' worker/public/data/*.json → 495 ×3, redis 135, sqlite 32; catalog of five in dashboard order; C1 groups list no VarMQDistributed; sqlite groups put the library on the x axis with ckpt1000/ckpt10000 as the series.

🤖 Generated with Claude Code

@fahimfaisaal
fahimfaisaal marked this pull request as ready for review October 24, 2026 23:00
fahimfaisaal and others added 9 commits October 25, 2026 08:35
Every subject opens SQLite with sqliteq's connection settings (WAL,
synchronous NORMAL, 5 s busy timeout) and the row's wal_autocheckpoint;
pool size and everything else follow each library's own defaults and
documentation. Rows are now <scenario>/ckpt<n>/<subject>, so the compared
library sits on the x axis like the pool and redis charts.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The default 100 ms FetchCooldown caps 16 workers at 160 jobs/s whatever
SQLite does (measured 159 at both checkpoint sizes). At the 1 ms minimum
the row is bound by SQLite and River's own per-job work. Listed as the
suite's one forced setting.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
rmq sleeps pollDuration after every fetch pass, fetched or empty, so the
example value capped one connection near prefetch / 0.1 s: the first run
measured 5.9k to 10.3k jobs/s at C1, on the cap, and a 102 ms pickup
latency that was the timer. rmq has no default for the setting; at 1 ms
the pass's own round trips set the number.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
32 rows, -benchtime=2x, varmq 415ad73 through the replace directive,
sqliteq v0.3.0, goqite v0.4.0, backlite v0.8.0, River v0.49.0 at a 1 ms
fetch cooldown, ext4 over NVMe.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Root, pool, redis and sqlite READMEs cut to results, subjects with
versions, matrix, running. The redis per-library source notes move to
redis/NOTES.md; site and deploy detail to worker/README.md. The site
builds five datasets: Worker Pool (24/8/4 CPUs), Message Queue Redis
(C1 and C4 groups together) and Message Queue SQLite. The redis README's
numbers are from the 2026-10-24 run until the rerun lands.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
150 rows, -benchtime=2x, same pinned setup as the first run. rmq moves
from 5.9k to 10.3k jobs/s at C1 to 27k to 33k and leads every
single-producer Noop row; its pickup latency drops from 102 ms to 1.5 ms.
README tables and the root headline regenerated from the file.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- River's 1 ms FetchCooldown moves to the queue config, so the client-level
  value that also paces its insert notifications stays at the default.
- walMB is the largest total of the run's -wal files, so one file per queue
  is measured like every queue in one file; sampling starts before open.
- Drain runs build the VarMQ worker with WithAutoRun(false) and Start it,
  instead of a non-blocking Pause that could let a credit pass run jobs
  during the untimed fill.
- goqite's priority runs in the same order as varmq and River.
- A failed row closes its consumer; a failed add fails the row at once;
  backlite and River stop errors are returned; the leftover poll after the
  last job is every 10 ms.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
32 rows, -benchtime=2x. River at a 1 ms per-queue fetch cooldown, walMB
as the total of the run's WAL files, drain runs started from a
not-yet-running worker. sqlite README and the root headline regenerated
from the file.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@fahimfaisaal
fahimfaisaal merged commit 1398ce3 into main Oct 25, 2026
1 check passed
@fahimfaisaal
fahimfaisaal deleted the feat/sqlite-rivals branch October 25, 2026 02:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant