Clean up the deployment host
ClosedPublic
Actions

Authored by thcipriani on Dec 9 2016, 1:14 AM.

Details

Maniphest Tasks

T112509: scap3 should repack / pack-refs git repos under /srv/deployment

Reviewers

dduvall
mmodell
demon
hashar

Group Reviewers

Release-Engineering-Team

Commits

rMSCA38289c69e15a: Clean up the deployment host

Patch without arc

git checkout -b D503 && curl -L https://phabricator.wikimedia.org/D503?download=true | git apply

Summary

I've been watching beta since we've started committing to git for every
deployment, and it seems that the git directory gets out-of-hand pretty
quickly. Leaving loose objects laying around when we flatten the
/srv/mediawiki directory makes for a large /srv/mediawiki/.git
directory.

The flattening of the mediawiki directory has been running for a week on
beta. The size of /srv/mediawiki/.git yesterday was 829 MB, after a
git gc it was 334 MB.

This commit runs git gc --auto for every deployment on
/srv/mediawiki. It also handles clean up of tags for all scap3 repos,
since there are a lot hanging around.

Should fix T112509

Diff Detail

Repository

rMSCA Scap

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

thcipriani updated this revision to Diff 1340.Dec 9 2016, 1:14 AM

thcipriani retitled this revision from to Clean up the deployment host.

thcipriani updated this object.

thcipriani edited the test plan for this revision. (Show Details)

thcipriani added reviewers: demon, mmodell, dduvall, hashar.

Restricted Application added a reviewer: Release-Engineering-Team. · View Herald TranscriptDec 9 2016, 1:14 AM

Restricted Application added a project: Release-Engineering-Team. · View Herald Transcript

Couldn't we set this in ./.git/config for a given repo so it happens more frequently/aggressively?

On further thought...we probably want to run this frequently, but tune some of the settings so --auto does really useful things each time. Default retention periods for unreferenced stuff (and reflog) could be shorter.

Also I'm finding some rumblings of a gc.autodetach config that runs it in the background so we'd be non-blocking.

These two settings would probably help:

gc.auto = 1000
gc.pruneexpire = now

It seems that gc.autodetach defaults to true when it's available so there is no need to set it.

We might want to experiment with gc.autopacklimit since that could have a large impact on rsync efficiency (as well as git fetch efficiency once we switch to using git transport)

mmodell accepted this revision.Dec 9 2016, 7:18 AM

mmodell edited edge metadata.

This revision is now accepted and ready to land.Dec 9 2016, 7:18 AM

hashar added a task: T112509: scap3 should repack / pack-refs git repos under /srv/deployment.Dec 9 2016, 11:10 AM

hashar updated this object.

hashar edited edge metadata.

hashar updated the revert plan for this revision. (Show Details)