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
durable: an instance touched by two versions in a rolling deploy can't be recovered by either #253
#235 raises durable's journal format to 2, and from then on Recover refuses a journal whose Records go below a format it has already seen. An older binary also refuses a Record whose format is lower than its checkpoint. So in a rolling deploy, an instance a new node starts (format-2 baseline) and an old node then fires on (format-1 Record) can't be recovered by either version:
the new version reports that an older version appended to a newer journal;
the old version reports that the Store didn't keep Record.Format.
Nothing documents a way to get the instance back, short of editing the journal by hand.
#235 documents "don't run mixed versions on one instance", and binaries already built can't be changed. The question is what recovery path to offer an operator who hits it anyway.
Options
a recovery tool, or a Recover option, that truncates the tail at the first Record from the other version, keeping the steps up to it;
deploy guidance: drain or pin instances to one version during a rollout;
a format-compatibility window, where version N also writes format N-1 until a flag flips.
Activity
fix(durable,state): a timer transition drives services and actors; st…