We deployed D16714 and reran one of the failing tasks -- it completed cleanly as expected. Thanks!

Yep! It looks like that's the issue:

I'll give that a try. I think the associated diffs have enormous lines, so it seems likely that's what's going on.

(Following up for Greg, as I'm now the oncall.)

mysql> select @@max_allowed_packet;
+----------------------+
| @@max_allowed_packet |
+----------------------+
|             67108864 |
+----------------------+
1 row in set (0.00 sec)

metamta.differential.inline-patches is 500.
metamta.email-body-limit is still the default of 524288.

Thanks for the quick feedback and for spending the time to help me do things The Right Way (instead of the okay but expedient way)!

"overwrite" + "update" sounds good to me! Changed.

With regards to the previous names: I think you got them backwards from what we originally came up with. Actually, they may be significantly better that way. Hmm...

The name change was motivated because I realized that the two operations differed only by level of granularity. It makes more sense to me if I imagine a set_commit level as well:

make (branch, commit, file) entries unique in repository_coverage; tweak mode names

Actually, would it make sense to enforce this uniqueness at the SQL level?

Oops! I absolutely agree the database should have at most one row per <branchID, commitID, pathID> tuple. I was confused about the behavior of the insert. Should've checked the database. <_<

Conspicuously absent is the "merge" mode we discussed because I've forgotten how to write PHP. If you want that in this diff I can add it, though

To bring this a bit back on topic, I think adding a Remarkup-rendered view that was not diff aware (as a first pass) would be very useful at pretty low implementation cost.

It looks like adding GFM or other markdown renderers has negative security implications, among other issues (see this discussion and T2849). We just want some sort of markdown renderer (i.e. remarkup) for checked-in files, which I think (without too much investigation) should be pretty straightforward to implement.