perf(thornode): console log writer looks up the working directory once
🔎 Visual review ofb20a294c: explainer, review flags and prompts to ask an AI about the change
Summary
thornode logs through the SDK's zerolog console logger, with the caller added by cmd/thornode (CallerWithSkipFrameCount(3)). For every line that has a caller, zerolog's default console caller formatter calls os.Getwd() (which stats . and $PWD) and then filepath.Rel to shorten the path. Release builds use -trimpath, so the caller is a module path such as gitlab.com/thorchain/thornode/v3/x/…. Rel can't relate that to the working directory, so the path never gets shortened, and the syscall is wasted on every line. Block execution logs a lot (every swap, every EndBlock manager), and the console writer runs on the block thread.
The start command now writes the SDK's logger through the same zerolog.ConsoleWriter with a caller formatter that looks the working directory up once. Everything else is kept as is:
zerolog.Logger.Outputkeeps the logger's level, hooks (timestamp), stack setting and context.- The writer gets the same
NoColor,TimeFormat(time.Kitchen) and output (cmd.OutOrStdout()) thatserver.CreateSDKLogger/cosmossdk.io/log.NewLoggeruse. - The formatter is zerolog's own, with the
os.Getwdcall moved out. - JSON output (
--log_format json) and module-filtered levels (e.g.consensus:info,*:error) go through SDK writers this can't rebuild, so they keep the SDK logger unchanged.
Risk: low. Log output only; no state. The output is byte-identical to the SDK logger's; see Tests.
What changes
cmd/thornode/console_log.go(new):consoleLogOutputplusconsoleFormatCaller, which is zerolog'sconsoleDefaultFormatCallerwith the working directory cached, andconsoleColorize, a copy of zerolog's unexportedcolorize.cmd/thornode/commands.go:addModuleInitFlags'PreRunEcalls it before wrapping the logger inthorlog.SdkLogWrapper.
Evidence
Benchmark (BenchmarkConsoleLogCaller, one error line with a -trimpath-style caller, 5 runs):
| ns/op | B/op | allocs/op | |
|---|---|---|---|
| SDK console writer | 13,013–15,166 | 3,001 | 65 |
| cached cwd | 9,398–10,525 | 2,520 | 61 |
That is about −32% per log line.
Mainnet replay (scripts/perf/mainnet-replay.sh, 1,999 blocks 28124001..28125999 through the CometBFT handshake, which checks every block's app hash against mainnet). This change was combined with !5283 (merged) and a scheduled-heights txout window on the v3.20.3-based replay lab stack, 3 interleaved rounds per side. Summed CPU profiles show this change's part:
| before | after | |
|---|---|---|
os.Getwd |
4.06 s | not in profile |
zerolog.ConsoleWriter.Write |
20.2 s | 17.4 s (−14%) |
That is about 0.8% of FinalizeBlock CPU (536.8 s).
scripts/perf/guard.sh (unit tests, then unfiltered regtest vs develop): PASS. All 2,481 regression output files are identical (state, events, gas_used).
Tests
cmd/thornode/console_log_test.go:
TestConsoleLogOutputMatchesSDK: the logger is built withserver.CreateSDKLoggerand withconsoleLogOutput, withNO_COLORset and unset, for default,info,debugwithlog_no_color, anderrorwith plain format. The same lines are logged through both: callers inside the working directory (shortened), a-trimpathmodule path, error fields, a module field, and no caller. Their output must be byte-identical. If a later SDK or zerolog upgrade changes the console writer, this test fails.TestConsoleLogOutputKeepsTrace: with--tracethe error stack is still logged.TestConsoleLogOutputKeepsOtherWriters: JSON output and module-filtered levels return the SDK logger unchanged.
How it was found
This came from round 8 of the mainnet replay profiling (rounds 4–7: !5274 (merged) to !5281 (merged)). Round 7 (!5281 (merged)) cut the error-log volume. This removes the per-line syscall that remained.