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
Commit 8018a0d
Browse filesBrowse the repository at this point in the historyBrowse files
environmentd: keep a restarted read-only process from booting writable
Address review on the in-place restart.
The entrypoint now exports `MZ_RESTARTED_IN_PLACE=true` before it
restarts environmentd, and `preflight_0dt` halts instead of booting
writable when that is set and the catalog is at the process's own
generation. A process restarts in place only while it is read-only and
has not announced itself ready to promote, so that combination means
another process of its generation, in a pod that replaced this one, has
been promoted. Booting writable would fence it. Without the marker, a
stale process could restart while the replacement was being promoted
and fence it once.
With that guard, the storage controller's read-only halt for a
collection deleted during bootstrap uses `restart_in_place!` too, and
keeps `halt!` when writable.
`restart_in_place!` exits with 75 (`EX_TEMPFAIL`) instead of 168, which
is also what a shell reports for a process killed by signal 40.
Restore `tini -g` and the entrypoint's TERM/INT traps, so environmentd
receives SIGTERM itself and the shell exits instead of restarting or
sleeping after it.
Shorten the halt in `check_ddl_changes` once the generation has been
promoted to "deployment generation N has been promoted; halting".
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
0 commit comments