The computed limits (limitFactor 3 x request = 1536Mi memory, 30Mi ephemeral
storage) were too small for real documents: a 12 MB HTML already gets the
container OOM-killed, and the caller sees no HTTP error at all, just a dropped
connection and a rising RESTARTS counter.
Rather than raise the requests and reserve conversion-sized capacity on the
node around the clock, requests and limits are decoupled: the requests stay as
they were, and both limits are now set explicitly. The chart already supported
this - an explicit limit wins over the limitFactor computation.
Measured in a throwaway k3d cluster with the official image (HTML via
Chromium), building on the series in the issue:
18 MB HTML, 2Gi -> OOMKilled
18 MB HTML, 2560Mi -> OOMKilled
18 MB HTML, 3Gi -> HTTP 200 in ~22s, 21 MB PDF
12 MB HTML, 1536Mi -> OOMKilled (the old computed limit)
12 MB HTML, 2Gi -> HTTP 200 in ~13s
Memory beyond 3Gi buys nothing: gotenberg's own request timeout is 30s and the
chart passes no flags to change it, so a document needing more memory has
already run out of time (a 25 MB HTML misses the 30s even at 4Gi).
The ephemeral-storage limit is raised in the same change, because raising
memory alone only moves the failure. With 3Gi memory and the old 30Mi limit,
/tmp peaked at 40 MB on a single conversion and at ~125 MB over a short
series, and the pod was evicted:
Pod ephemeral local storage usage exceeds the total limit of containers 30Mi
That failure is worse than the OOM-kill: the pod does not restart, it is
replaced, and the old one is left behind as Evicted.
The defaults in the effectiveResources helper are updated in lockstep with
values.yaml, so an erased "resources:" block (#99) falls back to the new
explicit limits instead of the old computed ones.
NOTES.txt is pulled along. Its numbers came from effectiveResources already,
but the text described a threshold that has moved and advice that no longer
holds: raising the request no longer drags the limit with it now that the
limits are explicit. The warning is kept but no longer claims memory is the
first limit you hit - with these limits, the 30s timeout is.
Major, because the runtime behaviour changes and the pod may occupy more
memory during large conversions than it was allowed to before.
Fixes#106
jklein
0c403b6100