You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Building Kex from source with unbounded --parallel OOM-kills on standard GitHub-hosted runners #363
Downstream repos (rodolfo, marqraft-cli, markdown) build Kex+Tey from main in CI, since no release yet has the manifest/stdlib features they need (documented in rodolfo's .github/workflows/ci.yml). That build does:
On a standard ubuntu-latest runner (2 vCPU, 7GB RAM), --parallel (unbounded — one cc1plus per core) now reliably OOM-kills partway through:
c++: fatal error: Killed signal terminated program cc1plus
compilation terminated.
gmake[2]: *** [.../lspgen.cpp.o] Terminated
Confirmed twice on marqraft-cli's CI: once killed after ~4m44s, on
7D16
ce (after a restart, presumably a heavier commit) after ~17m of apparent thrashing before dying. Capping to --parallel 1 on marqraft-cli's side made the build succeed, at the cost of a much longer build.
This wasn't a problem as of rodolfo's last few CI runs (all green in ~4 minutes, same --parallel, same runner class) — the build has apparently grown enough since then (now also compiling the LSP framework's lspgen and ket_template_test alongside the compiler itself) to no longer fit two parallel heavy C++ translation units in 7GB.
Not asking for anything specific here — mainly flagging it so downstream CI doesn't quietly start red across the org as the build keeps growing, and in case there's an obvious lever on this side (e.g. a lighter default target for consumers who only need kex+tey, not the LSP server / test binaries).
Already addressed on main — see README.md's "Building from source in CI" section and the JOBS/CMAKE_BUILD_JOBS bounding in the Makefile (commit 57b435b, which cites this issue by number). make build now bounds parallelism to the host's CPU count by default, and the README explains the unbounded---parallel OOM risk directly plus recommends cmake --build build --target kex --parallel 2 (an explicit, conservative job count — not bare --parallel) for small downstream CI runners, noting that even --target kex still builds the vendored LSP framework so isn't free on its own.
No further lever needed on this side beyond what's already documented; downstream repos (rodolfo, marqraft-cli, markdown) adopting the documented invocation should resolve their OOM kills.
Activity
akoskovacs commentedon Sep 20, 2026
Already addressed on
main— seeREADME.md's "Building from source in CI" section and theJOBS/CMAKE_BUILD_JOBSbounding in theMakefile(commit 57b435b, which cites this issue by number).make buildnow bounds parallelism to the host's CPU count by default, and the README explains the unbounded---parallelOOM risk directly plus recommendscmake --build build --target kex --parallel 2(an explicit, conservative job count — not bare--parallel) for small downstream CI runners, noting that even--target kexstill builds the vendored LSP framework so isn't free on its own.No further lever needed on this side beyond what's already documented; downstream repos (rodolfo, marqraft-cli, markdown) adopting the documented invocation should resolve their OOM kills.
Closing.