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
New threads did not start on an install whose provider container had become slow. A spawn that takes 1 s took the dispatch program more than 90 s, Amp timed out after 300 s and sent the spawn again with a new token, and the orb never got its setup requests. The cause of the slow container is not known: it had run for 7 days, and a new lite container handled 10 spawns at once and 200 jobs in 15 min in tests. Three changes make the provider safe against it:
Watchdog: every 5 min the Worker asks the provider container to run true. If it does not answer in 5 s two times in a row, the Worker restarts the container. No check in the first 10 min after a start, and at most one such restart in 30 min.
dispatch.sh starts 2 processes for a job that is already done (was about 25). It reads the Worker's new plain-text replies (?format=text) with shell builtins and needs only sh, curl and sleep (no awk or sed).
The Worker sends progress lines in groups, so dispatch polls less often.
/v1/status no longer waits on a provider container that does not answer. The log read stops after 5 s and the log tail says so. Before, one slow read blocked cf-orbs status, cf-orbs doctor, cf-orbs logs --provider and the 5-min provider check.
Added
Config providerInstance: size of the provider container, "lite" (default), "standard-1" or "standard-2". cf-orbs deploy restarts a running provider when the size changes.
cf-orbs doctor fails the provider check when the provider container runs but does not answer a log read.
GET /v1/ops/<id> and POST /v1/orbs/<orb>/ops accept ?format=text. The JSON replies are unchanged, so an older dispatch.sh on an external host keeps working.
Changed
External provider hosts: copy the new dispatch/dispatch.sh after you deploy the Worker. The new script needs a Worker from 0.2.5 or later.