batches: getting changed files in step: applying cached diff: error: unrecognized input: exit status 128

[sfllaw]

Sometimes, when updating a batch change that has already been applied, the cached diff cannot be applied:

sfllaw@hostname:~$ src batch apply -v -f batch-changes.yaml 
✅ Parsing batch spec
✅ Resolving namespace
✅ Preparing container images  ██████████████████████████████████████████████
🚧 Workspace creator: bind
✅ Set workspace type
✅ Resolved 137 repositories
✅ Found 137 workspaces with steps to execute
✅ Found 0 cached changeset specs; 137 tasks need to be executed

✅ Executing... (137/137, 24 errored)  ██████████████████████████████████████

❌ 24 errors:
   github.com/sfllaw/repo:
   getting changed files in step: applying cached diff: 'git apply -p0 /home/sfllaw/.cache/sourcegraph/batch/bind-workspace-test-1611338046' failed: error: unrecognized input: exit status 128
   Log: /tmp/changeset-github.com-sfllaw-repo-5e3e8f2c3e2780ec35240790d1971b2c0b7ed9b0.3325954318.log

If we look into the cache, we see that the diff is empty!

sfllaw@hostname:~/.cache/sourcegraph/batch$ jq -C . < github.com-sfllaw-repo-5e3e8f2c3e2780ec35240790d1971b2c0b7ed9b0/j7ViTNOX3Gkk_8XnX-MbZQ.json
{
  "diff": "",
  "changedFiles": {
    "modified": null,
    "added": null,
    "deleted": null,
    "renamed": null
  },
  "outputs": {},
  "path": ""
}

This is exactly the behaviour you’d expect from an empty patch:

sfllaw@hostname:~/src/repo$ echo -n '' > patch
sfllaw@hostname:~/src/repo$ git apply -p0 patch
error: unrecognized input
sfllaw@hostname:~/src/repo$ echo $?
128

[sfllaw]

cc: @sourcegraph/batchers (@eseliger @LawnGnome @courier-new @adeola-ak @BolajiOlajide @Piszmog @malomarrec @chrispine @danielmarquespt)

[BolajiOlajide]

Hello @sfllaw,

Thanks for reporting this issue.
We've added it to our planning discussion list and we will respond after.

Cheers!

[LawnGnome]

@sfllaw At the risk of causing this to take much longer than it should, does using the -clear-cache flag "fix" the issue? (We're just trying to narrow down where the missing check is, since we should handle empty diffs more gracefully.)

[sfllaw]

@LawnGnome I’ve finished working on this batch change, so I’m unable answer that question. But I’m guessing the answer is yes?

I recommend that an empty cache diff should not create a changeset because there is no change? Skipping seems to be the what happens on a fresh change, so I’d expect the same behaviour on subsequent runs.

[eseliger]

Found the cause, will fix this in the next release. Thanks for reporting and sorry for the inconvenience :)