Skip to content

ci(linux): fit the Devel Flatpak on the runner disk and fix two package builder test timeouts - #4295

Merged
datlechin merged 4 commits into
mainfrom
fix/flatpak-devel-runner-disk
Nov 15, 2026
Merged

datlechin merged 4 commits into
mainfrom
fix/flatpak-devel-runner-disk

Conversation

@datlechin

@datlechin datlechin commented Nov 15, 2026 •

Copy link
Copy Markdown
Member

Summary

  • Devel Flatpak manifest: CARGO_INCREMENTAL=0 and CARGO_PROFILE_DEV_DEBUG=line-tables-only in the build options.
  • Cause: the dev-profile build of the workspace and every test binary filled the runner disk. Probe run 43715567785 measured the build directory at 73 GB (debug/deps 52 GB, debug/incremental 18 GB) against 76 GB free on Azure and 79 GB on AWS. 4 of 5 probe jobs ran out, on main and on b451ff2, before fix(core): bound what a SQLite or DuckDB catalog read holds #4287; the one pass bottomed at 1.9 GB free. The nightly of 2026-11-14 failed the same way.
  • The tests are not the cause: the whole Meson test run wrote under 0.5 GB.
  • flatpak-builder starts each build in an empty build directory, so incremental output is never read again.
  • Line tables match the release manifest and the CI jobs: a backtrace keeps its files and lines.
  • DuckDB catalog_memory test: the crafted child has 60 keys and a 640 KiB name instead of 600 keys and 64 KiB. Each key DuckDB adds to the parent copies every earlier one, so the open cost keys squared times name, and it passed the 10 s connect timeout on the package builders (Debian resolute here, Debian stonking and Fedora 44 on main af5869a). The reads it checks cost keys times name, the same 39 MB as before.
  • CLI a_tool_call_the_app_never_answers_stops_at_its_deadline: the handshake, which the test server answers itself, gets the usual 15 s bound and only the tool call keeps 300 ms. On the Fedora 45 package builder the handshake took over 300 ms and the test failed in setup (NoAnswer { seconds: 0 } is 300 ms rounded down).

Test plan

  • Probe run 43722351617, Azure runners with 77 GB free: with this manifest the build directory peaked at 22 GB and the job at 25 GB in 3 of 3 jobs, every Meson suite passing. Incremental output off alone: 51 GB build directory, 55 GB job. The exported app went from 1.9 GB to 0.7 GB.
  • linux / Flatpak Devel (x86_64) on this PR, run three times.
  • catalog_memory locally: the file opens in 0.7 s against 1.8 to 2.3 s and is created in 0.9 s against 5 to 6 s; the test takes 2.5 s against 8.2 s. With the constraint estimate disabled it still fails, holding 64 MB for the keys into named_parent.
  • Debian package jobs on 33dfa88: catalog_memory takes 5.5 and 6.2 s against 17.7 to 24 s before.
  • verify.sh core tablepro-cli client::tests: 4 of 4 pass.

@datlechin
datlechin enabled auto-merge (squash) November 15, 2026 14:08
@datlechin datlechin changed the title ci(linux): build the Devel Flatpak with no incremental artifacts and line tables only ci(linux): fit the Devel Flatpak on the runner disk and open the crafted DuckDB catalog within the connect timeout Nov 15, 2026
@datlechin datlechin changed the title ci(linux): fit the Devel Flatpak on the runner disk and open the crafted DuckDB catalog within the connect timeout ci(linux): fit the Devel Flatpak on the runner disk and fix two package builder test timeouts Nov 15, 2026
@datlechin
datlechin merged commit ee56563 into main Nov 15, 2026
52 of 89 checks passed
@datlechin
datlechin deleted the fix/flatpak-devel-runner-disk branch November 17, 2026 04:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant