Add an optional benchmark workflow

[pid1]

Adds .github/workflows/benchmark.yaml, scripts/benchmark.sh and scripts/benchmark.lua. The shell builds two refs with the repo Dockerfile, starts a container for each and runs the interleaved loop; the Lua seeds redis, drives the load, reads INFO commandstats and writes the job summary.

Advisory. workflow_dispatch only, so it never runs on a push or a pull request and cannot turn anything red. It exits non-zero only when a build fails, a container never answers /healthcheck, or more than 1% of requests fail, never on a timing.

The driver runs inside the server container under /opt/openresty/bin/resty, which the image already installs. HTTP and redis are then both on loopback, so there is no container networking to arrange and redis needs no exposed port. The driver is copied in at run time, so it also works against a ref that predates it.

Method: arms interleaved across repetitions, FLUSHDB and a seed write before every measurement, a discarded warmup, CONFIG RESETSTAT immediately before each measured window, and every figure a paired ratio against the base within the same repetition. Redis commands per request are reported separately from latency and throughput, because the first is exact and the second is not trustworthy on shared hardware.

Run it from Actions, Benchmark, Run workflow. Locally: ./scripts/benchmark.sh --engine podman --base origin/master --head HEAD.

Verified under podman on an M3 Pro, d941a63 against current master, 3 repetitions of 5 seconds at 2 connections:

| endpoint                 | base | head | change |
| PUT /syncs/progress      | 6.00 | 5.00 |  -1.00 |
| GET /syncs/progress/C1   | 4.00 | 3.00 |  -1.00 |

PUT p50 254us -> 193us, -24.9% paired (-25.1..-23.1), rps +32.4% (+30.5..+32.9)
GET p50 243us -> 194us, -21.0% paired (-22.1..-17.7), rps +27.0% (+19.9..+27.0)
555395 requests, 0 non-200 or failed.

The per-command table attributes all of it to select, which is #58. Base and head both on master gives the noise floor on that machine: paired p50 within 7.2% and rps within 5.8%.

actionlint is clean; shellcheck reports only SC2329 on the trap handler.

The workflow has not run on a GitHub runner, so the timeout, the two builds against a cold cache, and whether the timing half is legible on two shared vCPUs are guesses. The command counts do not depend on any of that.

No new dependencies: everything the driver uses is in the image already.

[Frenzie]

Doing it in Python seems amusing. ;-)

It's potentially worth keeping in mind that GHA could pull https://github.com/koreader/koreader-sync-server/pkgs/container/kosync which I think contains the necessary Lua equivalents. It could even query Redis with redis-lua. But the networking might be annoying.

tl;dr Since you're using Claude extensively here see to what extent it thinks the suggestion is feasible. Not an objection to Python, mainly curiosity.

[pid1]

I'm happy to do whatever language. Python was by my direction, as I wasn't sure if this was something y'all would even be interested in. I can do a Lua rewrite, if that's preferable.

[Frenzie]

It's possible it'd be more awkward to do the same thing with Lua; it's just that I think it's the obvious first choice to try in a Lua repo. ;-) I'd say give it a go since it'll just take the LLM some five minutes to figure it out regardless; if it's too awkward we'll go with the Python.

[pid1]

Swapped it over to Lua. The driver runs inside the container with resty, so HTTP and redis are both on loopback and there's no networking to arrange.

It's a bit bigger than the Python, 571 lines against 423, mostly because the image has no HTTP client so it writes the requests itself.

Worth noting the rewrite caught something the Python had wrong: nginx closing a keepalive connection after keepalive_requests was being counted as a failed request.

[Frenzie]

It's a bit bigger than the Python, 571 lines against 423, mostly because the image has no HTTP client so it writes the requests itself.

It probably doesn't matter too much, but it includes LuaSocket which definitely comes with one. Something like local http = require("socket.http").

[pid1]

You're right that it's in there. It's blocking though, so under resty it would serialise the ngx.thread coroutines and the timings would come out wrong. Cosockets are the nonblocking equivalent, which is why it ended up hand-rolled.