Repository navigation
exec.sh: ~69 s of bookkeeping plus a 165 s deploy-build gate for a 1 s model write (Windows profile) #585
Description
Per-command profile of one bin/exec.sh run on Windows (Git Bash 5.3 / Cygwin), toolkit c28fc0d, mxcli v0.28.0, Mendix 11.12.6. The run was an idempotent re-run of an already-applied 5-statement script: the model write itself took 1.1 s, and the whole run took 234 s.
How measured: export PS4='+${EPOCHREALTIME} ${BASH_SOURCE##*/}:${LINENO} ' SHELLOPTS; bash -x bin/exec.sh <script>. Each line is charged the gap until the next traced line (5,137 lines).
Where the time went
| Phase | Time | Main cost |
|---|---|---|
| guards | 8.9 s | exec-approval, mdl-scan, row-groups check 2.6 s |
| check | 9.0 s | mxcli check via tee 3.5 s, script-owners check 2.4 s, mxtk_model_version 1.4 s |
| snapshot | 7.2 s | cp -r mprcontents 2.7 s, rm -rf old snapshot 0.8 s |
| baseline | 7.2 s | model-stamp.sh check -q --clean 5.5 s |
| exec | 1.2 s | the actual write |
| gate | 165.8 s | one mxbuild --target=deploy: 165 s |
| lint | 6.5 s | lint-gate.sh 5.9 s |
| score | 5.7 s | metrics-ledger score 5.5 s |
| after verdict | 23.0 s | build-plan-status --refresh 8.3 s, metrics-ledger exec 3.0 s, script-owners record 1.7 s |
Totals: 165 s mxbuild, ~34 s before the write, ~35 s after the gate. About 69 s is toolkit bookkeeping for a 1 s write.
On this machine (corporate Windows 11, Defender real-time protection on): a $(...) subshell costs ~58 ms, an external program ~86 ms, a builtin ~0. The 4,972 traced lines that took under 0.5 s add up to 24.9 s, which is the per-fork cost spread across the helpers.
Findings / suggestions, by time saved
- The per-exec gate is a full deploy build. It took 165 s in this run, 18-20 s in warm runs earlier the same day, and ~156 s cold.
mx checkon the same model takes 12 s and reports the same 0 errors. Suggestmx checkas the per-exec gate, with the deploy build (javac etc.) only when Java/JS actions changed, at module close, or before commit. At minimum, print "full deploy build, can take 2-3 min" so the run doesn't look hung. - 23 s of work after the verdict is already known. build-plan-status
--refresh(8.3 s), the ledger row and script-ownersrecordcould run in the background or be deferred to commit/close-task. - lint (5.9 s) and score (5.5 s) are two separate passes over the same model. One pass could produce both.
- Repeated model and stamp reads before the write.
model-stamp check5.5 s,mxtk_model_version1.4 s (still called whenMXBUILD_PATHis set), row-groups 2.6 s, script-owners 2.4 s. Memoize the model version and stamp per run (keyed on the .mpr mtime), and fold the two script-owners calls into one. - Snapshot copies the whole
mprcontents/(2.7 s, and growing with the model). Considergit stash/a commit-based snapshot, or hardlinks, when the tree is clean. - Fork-heavy helpers. On Windows every
$(...),| sed,| head,dirnameordatecosts 60-90 ms. Prefer builtins (${var%/*},printf -v,$EPOCHSECONDS) on the hot path. - Ship this profiler. An
MXTK_PROFILE=1switch that sets the PS4 above and prints the per-phase table would let users file numbers like these without digging.
Complements the earlier field report "exec.sh performance on Windows (Git Bash): where the time went" (setting MXBUILD_PATH saves ~6 s/run; spawn cost is ~94 ms; 4 of 9 runs blocked on an uncommitted model).
Activity
MendixMau commentedon Oct 28, 2026
Triage: DISCUSS track. Performance, plus a process path (the per-exec gate) in every entry mode that runs
exec.sh(migration, requirements-driven, greenfield, change an existing app). Code paths verified; timings UNVERIFIED (Windows profile not reproduced here).Problem:
project-bin/exec.sh:1850→mxtk_mxbuild_error_count, which runsmxbuild --target=deploy(project-bin/_common.sh:910). Cold, that took 165 s against a 1.1 s write.mxtk_ensure_mxbuildcallsmxtk_model_versioneven whenMXBUILD_PATHis set (_common.sh:1050). Thenmodel-stamp.sh check(exec.sh:1567),row-groups.sh check(exec.sh:618) andscript-owners.sh check(exec.sh:1230) each read it again.build-plan-status --refreshruns synchronously (exec.sh:2234).MXTK_PROFILEanywhere on master.Value: High. This is the inner build loop. Your numbers put model work under 5% of each run.
Options:
Profiler first.
MXTK_PROFILE=1prints a per-phase table and counts spawns, for bothexec.shandgate-check.sh. This is the same switch as Stage 3 (Architecture & Design) takes ~1h50m wall time on a 1-module, 7-page demo slice; gate-check + closeout alone ~12 min #583 option 1: one instrument, two reports. No path changes, near-zero cost when off. Small.Cheap wins, no process change.
MXBUILD_PATHis set.script-ownerscalls into one.Estimated saving: about 15-25 s per run on Windows. Medium. The risk is in the background step: a ledger or status row could still be being written when the next exec starts.
Lighter per-exec gate: run
mx check(about 12 s) on every exec. Run the deploy build only when Java/JS actions changed, at module close, or before commit. Saves around 150 s cold and 6-8 s warm. This changes the process: a javac-level error is caught later, at module close instead of at the exec that caused it. Large, and it needs a field run on both layouts.Recommendation: 1 now, then 2. Decide on 3 once the profiler has numbers from at least one macOS project and one Windows project. 3 is the biggest saving, but it moves when build errors are caught.
Decision needed from you: build the shared profiler (option 1, which also settles #583), or go straight to option 2 or 3? Reply "go option N" to build it.
Generated by Claude Code
Exec speed: mx check gate by default, doctor performance check, Windo…
exec: ~60% fewer process starts per exec (Windows overhead, #590)
doctor --bench, machine health, recommended fix, and the speed-setup …
MendixMau commentedon Nov 3, 2026
Thanks @mkr87 — closing as fixed: the per-exec bookkeeping was trimmed (#596/#597) and
Gate mode: checknow gates each exec withmx checkinstead of a full deploy build; the full build runs at the module boundary (bin/module-close.sh). Please reopen with a fresh timing if you still see it on current master.Generated by Claude Code