shoney/writeonce
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

perf(db): group commit measured — ~2.9x durable write throughput, T5

Browse source 
databasev2 4 part A, task 5.

Controlled before/after — same machine, same workload (wmix 4000 32),
same build except db.c and vm.c, two runs each interleaved:

- per-statement barrier: 2213 / 2177 ops/sec, p50 7183 / 7251us
- group commit:          6216 / 6525 ops/sec, p50 3458 / 3444us
- ~2.9x throughput, ~2.1x lower p50

The full campaign confirms it a second way: s1 takes the inline path and
commits per statement BY DESIGN, so within one build the shard configs are
batching-off vs batching-on — 1467 -> 5117 ops/sec, mean batch 1.0 -> 5.43,
peak 1 -> 57. 3.5x, agreeing with the 2.9x above.

Recorded honestly:

- the BEFORE p99 is at the histogram ceiling (hist_add clamps at 20000us
  and both runs pinned there), so the true figure is >=20ms and unknown.
  The improvement is AT LEAST 2.3x; the old p99 was off the instrument
- durable.sN.mixwrite went 480 -> 492 ops/sec, i.e. UNCHANGED. That was
  the spec's original payoff metric and correcting it was part of the
  brainstorm: mix performs 20 writes at C=4, mean batch 1.01. A workload
  that never has two writes in flight cannot be helped by batching them
- seed is likewise unchanged: a serial writer has nothing to batch with
- so the payoff is real but CONDITIONAL — it appears where concurrent
  durable writes fan into the owner shard, and nowhere else

Two traps recorded in perf-targets §6:

- do not benchmark durability on /tmp: it is tmpfs here, where fdatasync
  is free. The same run reported 195000 ops/sec at p50 1us there against
  2200 at p50 7200us on ext4 — no barrier to amortise, so the measurement
  measures nothing. db-bench keeps its stores under bench/ for this reason
- the record count is not the update count: 7755 records for 4000 updates,
  because hist_dump and the done-marker are themselves durable inserts

- FIXED a regression I introduced in T4: master's committed baseline is
  FULL mode (N=20000, crash_reps=3, msg_n=200000) and I had overwritten it
  with quick-mode values. Regenerated from a full campaign; the full run
  now passes 106 checks 0 failures against it
- gate still bites: sN wmix ops_sec -70% -> FAIL on exactly that metric

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
shoney.arickathil 2026-08-28 09:54:52 +02:00
parent 5f9598af6a
commit 6183a67dfc
2 changed files with 197 additions and 126 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

252
bench/baseline.json
View file

@ -1,52 +1,52 @@
{ {
"_config": { "_config": {
"N": 2000, "N": 20000,
"crash_reps": 1, "crash_reps": 3,
"msg_n": 20000, "msg_n": 200000,
"note": "refresh only with a commit that says why; tolerances come from tolerance_for() in the driver", "note": "refresh only with a commit that says why; tolerances come from tolerance_for() in the driver",
"wal_n": 800 "wal_n": 4000
}, },
"durable.s1.mixread.ops_sec": { "durable.s1.mixread.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 2240, "floor": 2070,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 8961 "value": 8281
}, },
"durable.s1.mixread.p50us": { "durable.s1.mixread.p50us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 2
}, },
"durable.s1.mixread.p99us": { "durable.s1.mixread.p99us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 4 "value": 21
}, },
"durable.s1.mixwrite.ops_sec": { "durable.s1.mixwrite.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 248, "floor": 230,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 995 "value": 920
}, },
"durable.s1.mixwrite.p50us": { "durable.s1.mixwrite.p50us": {
"dir": "lower", "dir": "lower",
"floor": 824, "floor": 1784,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 206 "value": 446
}, },
"durable.s1.mixwrite.p99us": { "durable.s1.mixwrite.p99us": {
"dir": "lower", "dir": "lower",
"floor": 904, "floor": 2252,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 226 "value": 563
}, },
"durable.s1.query.ops_sec": { "durable.s1.query.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 233644, "floor": 210260,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 934579 "value": 841042
}, },
"durable.s1.query.p50us": { "durable.s1.query.p50us": {
"dir": "lower", "dir": "lower",
@ -58,13 +58,13 @@
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 3 "value": 4
}, },
"durable.s1.read.ops_sec": { "durable.s1.read.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 190114, "floor": 284155,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 760456 "value": 1136621
}, },
"durable.s1.read.p50us": { "durable.s1.read.p50us": {
"dir": "lower", "dir": "lower",
@ -80,21 +80,21 @@
}, },
"durable.s1.seed.ops_sec": { "durable.s1.seed.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 1080, "floor": 1076,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 4323 "value": 4306
}, },
"durable.s1.seed.p50us": { "durable.s1.seed.p50us": {
"dir": "lower", "dir": "lower",
"floor": 848, "floor": 868,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 212 "value": 217
}, },
"durable.s1.seed.p99us": { "durable.s1.seed.p99us": {
"dir": "lower", "dir": "lower",
"floor": 2108, "floor": 2080,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 527 "value": 520
}, },
"durable.s1.wmix.mean_batch": { "durable.s1.wmix.mean_batch": {
"dir": "higher", "dir": "higher",
@ -104,21 +104,21 @@
}, },
"durable.s1.wmix.ops_sec": { "durable.s1.wmix.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 810, "floor": 366,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 3241 "value": 1467
}, },
"durable.s1.wmix.p50us": { "durable.s1.wmix.p50us": {
"dir": "lower", "dir": "lower",
"floor": 852, "floor": 1820,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 213 "value": 455
}, },
"durable.s1.wmix.p99us": { "durable.s1.wmix.p99us": {
"dir": "lower", "dir": "lower",
"floor": 2032, "floor": 2884,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 508 "value": 721
}, },
"durable.s1.wmix.peak_batch": { "durable.s1.wmix.peak_batch": {
"dir": "higher", "dir": "higher",
@ -134,63 +134,63 @@
}, },
"durable.s1.write.ops_sec": { "durable.s1.write.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 1103, "floor": 554,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 4415 "value": 2216
}, },
"durable.s1.write.p50us": { "durable.s1.write.p50us": {
"dir": "lower", "dir": "lower",
"floor": 856, "floor": 1828,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 214 "value": 457
}, },
"durable.s1.write.p99us": { "durable.s1.write.p99us": {
"dir": "lower", "dir": "lower",
"floor": 1932, "floor": 2880,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 483 "value": 720
}, },
"durable.sN.mixread.ops_sec": { "durable.sN.mixread.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 1111, "floor": 1107,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 4445 "value": 4429
}, },
"durable.sN.mixread.p50us": { "durable.sN.mixread.p50us": {
"dir": "lower", "dir": "lower",
"floor": 348, "floor": 308,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 87 "value": 77
}, },
"durable.sN.mixread.p99us": { "durable.sN.mixread.p99us": {
"dir": "lower", "dir": "lower",
"floor": 17540, "floor": 4172,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 4385 "value": 1043
}, },
"durable.sN.mixwrite.ops_sec": { "durable.sN.mixwrite.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 123, "floor": 123,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 493 "value": 492
}, },
"durable.sN.mixwrite.p50us": { "durable.sN.mixwrite.p50us": {
"dir": "lower", "dir": "lower",
"floor": 1460, "floor": 2156,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 365 "value": 539
}, },
"durable.sN.mixwrite.p99us": { "durable.sN.mixwrite.p99us": {
"dir": "lower", "dir": "lower",
"floor": 3888, "floor": 6952,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 972 "value": 1738
}, },
"durable.sN.query.ops_sec": { "durable.sN.query.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 211864, "floor": 237529,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 847457 "value": 950118
}, },
"durable.sN.query.p50us": { "durable.sN.query.p50us": {
"dir": "lower", "dir": "lower",
@ -202,13 +202,13 @@
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 2 "value": 3
}, },
"durable.sN.read.ops_sec": { "durable.sN.read.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 204415, "floor": 276701,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 817661 "value": 1106806
}, },
"durable.sN.read.p50us": { "durable.sN.read.p50us": {
"dir": "lower", "dir": "lower",
@ -224,81 +224,81 @@
}, },
"durable.sN.seed.ops_sec": { "durable.sN.seed.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 1111, "floor": 1085,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 4446 "value": 4343
}, },
"durable.sN.seed.p50us": { "durable.sN.seed.p50us": {
"dir": "lower", "dir": "lower",
"floor": 848, "floor": 868,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 212 "value": 217
}, },
"durable.sN.seed.p99us": { "durable.sN.seed.p99us": {
"dir": "lower", "dir": "lower",
"floor": 1996, "floor": 2040,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 499 "value": 510
}, },
"durable.sN.wmix.mean_batch": { "durable.sN.wmix.mean_batch": {
"dir": "higher", "dir": "higher",
"floor": 0.0, "floor": 1.0,
"tolerance_pct": 100, "tolerance_pct": 100,
"value": 3.38 "value": 5.43
}, },
"durable.sN.wmix.ops_sec": { "durable.sN.wmix.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 1575, "floor": 1279,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 6302 "value": 5117
}, },
"durable.sN.wmix.p50us": { "durable.sN.wmix.p50us": {
"dir": "lower", "dir": "lower",
"floor": 13776, "floor": 32832,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 3444 "value": 8208
}, },
"durable.sN.wmix.p99us": { "durable.sN.wmix.p99us": {
"dir": "lower", "dir": "lower",
"floor": 51212, "floor": 48676,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 12803 "value": 12169
}, },
"durable.sN.wmix.peak_batch": { "durable.sN.wmix.peak_batch": {
"dir": "higher", "dir": "higher",
"floor": 7, "floor": 14,
"tolerance_pct": 100, "tolerance_pct": 100,
"value": 28 "value": 57
}, },
"durable.sN.wmix.peak_staged": { "durable.sN.wmix.peak_staged": {
"dir": "lower", "dir": "lower",
"floor": 5488, "floor": 11172,
"tolerance_pct": 100, "tolerance_pct": 100,
"value": 1372 "value": 2793
}, },
"durable.sN.write.ops_sec": { "durable.sN.write.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 1117, "floor": 552,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 4470 "value": 2210
}, },
"durable.sN.write.p50us": { "durable.sN.write.p50us": {
"dir": "lower", "dir": "lower",
"floor": 852, "floor": 1828,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 213 "value": 457
}, },
"durable.sN.write.p99us": { "durable.sN.write.p99us": {
"dir": "lower", "dir": "lower",
"floor": 2020, "floor": 2948,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 505 "value": 737
}, },
"ram.s1.mixread.ops_sec": { "ram.s1.mixread.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 2235, "floor": 22322,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 8940 "value": 89290
}, },
"ram.s1.mixread.p50us": { "ram.s1.mixread.p50us": {
"dir": "lower", "dir": "lower",
@ -314,9 +314,9 @@
}, },
"ram.s1.mixwrite.ops_sec": { "ram.s1.mixwrite.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 248, "floor": 2480,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 993 "value": 9921
}, },
"ram.s1.mixwrite.p50us": { "ram.s1.mixwrite.p50us": {
"dir": "lower", "dir": "lower",
@ -328,19 +328,19 @@
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 2
}, },
"ram.s1.msgrate.msgs_sec": { "ram.s1.msgrate.msgs_sec": {
"dir": "higher", "dir": "higher",
"floor": 384911, "floor": 1568873,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 3079291 "value": 12550988
}, },
"ram.s1.query.ops_sec": { "ram.s1.query.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 274725, "floor": 256016,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1098901 "value": 1024065
}, },
"ram.s1.query.p50us": { "ram.s1.query.p50us": {
"dir": "lower", "dir": "lower",
@ -352,13 +352,13 @@
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 2
}, },
"ram.s1.read.ops_sec": { "ram.s1.read.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 312500, "floor": 233448,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1250000 "value": 933794
}, },
"ram.s1.read.p50us": { "ram.s1.read.p50us": {
"dir": "lower", "dir": "lower",
@ -370,91 +370,91 @@
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 2
}, },
"ram.s1.seed.ops_sec": { "ram.s1.seed.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 335570, "floor": 53529,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 1342281 "value": 214119
}, },
"ram.s1.seed.p50us": { "ram.s1.seed.p50us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 1 "value": 4
}, },
"ram.s1.seed.p99us": { "ram.s1.seed.p99us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 2 "value": 12
}, },
"ram.s1.write.ops_sec": { "ram.s1.write.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 237416, "floor": 48021,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 949667 "value": 192086
}, },
"ram.s1.write.p50us": { "ram.s1.write.p50us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 1 "value": 6
}, },
"ram.s1.write.p99us": { "ram.s1.write.p99us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 15, "tolerance_pct": 15,
"value": 2 "value": 15
}, },
"ram.sN.mixread.ops_sec": { "ram.sN.mixread.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 2243, "floor": 7481,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 8975 "value": 29924
}, },
"ram.sN.mixread.p50us": { "ram.sN.mixread.p50us": {
"dir": "lower", "dir": "lower",
"floor": 264, "floor": 320,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 66 "value": 80
}, },
"ram.sN.mixread.p99us": { "ram.sN.mixread.p99us": {
"dir": "lower", "dir": "lower",
"floor": 2080, "floor": 624,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 520 "value": 156
}, },
"ram.sN.mixwrite.ops_sec": { "ram.sN.mixwrite.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 249, "floor": 831,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 997 "value": 3324
}, },
"ram.sN.mixwrite.p50us": { "ram.sN.mixwrite.p50us": {
"dir": "lower", "dir": "lower",
"floor": 288, "floor": 348,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 72 "value": 87
}, },
"ram.sN.mixwrite.p99us": { "ram.sN.mixwrite.p99us": {
"dir": "lower", "dir": "lower",
"floor": 344, "floor": 608,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 86 "value": 152
}, },
"ram.sN.msgrate.msgs_sec": { "ram.sN.msgrate.msgs_sec": {
"dir": "higher", "dir": "higher",
"floor": 176056, "floor": 248897,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1408450 "value": 1991179
}, },
"ram.sN.query.ops_sec": { "ram.sN.query.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 308641, "floor": 227790,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1234567 "value": 911161
}, },
"ram.sN.query.p50us": { "ram.sN.query.p50us": {
"dir": "lower", "dir": "lower",
@ -466,13 +466,13 @@
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 2
}, },
"ram.sN.read.ops_sec": { "ram.sN.read.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 310945, "floor": 216919,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1243781 "value": 867678
}, },
"ram.sN.read.p50us": { "ram.sN.read.p50us": {
"dir": "lower", "dir": "lower",
@ -484,42 +484,42 @@
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 2
}, },
"ram.sN.seed.ops_sec": { "ram.sN.seed.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 363372, "floor": 60469,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1453488 "value": 241878
}, },
"ram.sN.seed.p50us": { "ram.sN.seed.p50us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 4
}, },
"ram.sN.seed.p99us": { "ram.sN.seed.p99us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 2 "value": 10
}, },
"ram.sN.write.ops_sec": { "ram.sN.write.ops_sec": {
"dir": "higher", "dir": "higher",
"floor": 202922, "floor": 41677,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 811688 "value": 166708
}, },
"ram.sN.write.p50us": { "ram.sN.write.p50us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 1 "value": 8
}, },
"ram.sN.write.p99us": { "ram.sN.write.p99us": {
"dir": "lower", "dir": "lower",
"floor": 100, "floor": 100,
"tolerance_pct": 50, "tolerance_pct": 50,
"value": 4 "value": 15
} }
} }

71
docs/plan/perf-targets.md
View file

@ -67,3 +67,74 @@ not the limiting factor for any current workload).
the ~55× gap is one fdatasync per statement (~220µs each). the ~55× gap is one fdatasync per statement (~220µs each).
**Owner: iteration 23** (io_uring group-commit) — its acceptance is **Owner: iteration 23** (io_uring group-commit) — its acceptance is
literally this number moving while the crash battery stays green. literally this number moving while the crash battery stays green.
## 6. WAL group commit: one barrier per drain (databasev2 4 part A)
**Measured 2026-08-28.** Before this, the engine committed per *statement*:
`db.c` called `wo_wal_commit` immediately after every append, so each row
change bought its own `pwrite` + `fdatasync`. Now shard 0 stages every queued
write request, issues one barrier, and only then releases the held replies.
### The controlled before/after
Same machine, same workload (`wmix 4000 32` — every op a durable update, 32
concurrent), same build except `db.c` and `vm.c`, two runs each, interleaved:
| | ops/sec | p50 | p99 |
| --- | --- | --- | --- |
| per-statement barrier | 2213 · 2177 | 7183 · 7251 µs | **20000 · 20000 µs** |
| group commit | **6216 · 6525** | **3458 · 3444 µs** | 11139 · 5971 µs |
**≈2.9× throughput, ≈2.1× lower p50.**
**The p99 "before" figure is at the histogram ceiling, not a measurement.**
`hist_add` clamps at 20000 µs, and both before-runs pinned there — so the true
before p99 is ≥20 ms and unknown. The improvement is *at least* 2.3×; the
honest statement is that the old p99 was off the end of the instrument.
### Confirmation from the committed baseline
The full campaign gives the same answer a second way. `s1` takes the inline
path, which commits per statement **by design**, so within one build the two
shard configurations are batching-off against batching-on:
| Leg | ops/sec | p50 | p99 | mean batch | peak batch |
| --- | --- | --- | --- | --- | --- |
| `durable.s1.wmix` (inline, unbatched) | 1467 | 455 µs | 721 µs | **1.0** | 1 |
| `durable.sN.wmix` (batched) | **5117** | 8208 µs | 12169 µs | **5.43** | 57 |
3.5× throughput, agreeing with the 2.9× above. Note `sN` latency is *higher*
while throughput is 3.5× better: 64 writers queueing behind one owner shard
trade per-op latency for barrier amortisation, which is what group commit is.
Batching scales with write concurrency exactly as designed — mean batch at
C = 4 / 16 / 64 was **1.13 / 1.76 / 5.35**, peak **3 / 10 / 39**.
### What did NOT improve, and why that was predicted
`durable.sN.mixwrite` went **480 → 492 ops/s** — unchanged. That is the metric
the spec *originally* named as the payoff, and correcting it was part of the
brainstorm: `mix` writes on one op in ten with C=4, so a quick run performs
**20 writes** and mean batch measured **1.01** over 3112 barriers. A workload
that never has two writes in flight cannot be helped by batching them.
`durable.*.seed` is likewise unchanged: a serial single writer has nothing to
batch with under any scheme.
**So the payoff is real but conditional: it appears exactly where concurrent
durable writes fan into the owner shard, and nowhere else.**
### Two traps worth recording
**Do not benchmark durability on `/tmp`.** It is `tmpfs` here, where
`fdatasync` is free — the same `wmix` run reported **195 000 ops/s at p50 1 µs**
there against **2200 ops/s at p50 7200 µs** on ext4. There is no barrier to
amortise on a memory filesystem, so a group-commit measurement taken there
measures nothing. `db-bench` gets this right by keeping its stores under
`bench/`.
**The record count is not the update count.** `wmix` staged 7755 records for
4000 updates because the histogram dump and the done-marker are themselves
durable inserts. They arrive as an end-of-run burst, which is batch-friendly,
so `mean_batch` is not purely update-driven. Peak staged bytes stayed small
(2793 B at C=64), which is what settled the decision to ship **no batch cap**:
the request queue's existing upstream bound is sufficient.