It's possible to attach files in a cycle (A -> B -> A) and possibly to themselves (A -> A) by commenting on file A with {Fbbbb} and on file B with {Faaaa}. The policy evaluation never exits.
A hacky fix would be to prevent files from being attached to files.
A cleaner fix would be to resolve cycles correctly, but this might be involved.
Definitely my problem:
MariaDB [phabricator_file]> select * from edge order by dateCreated desc limit 5; +--------------------------------+------+--------------------------------+-------------+-----+--------+ | src | type | dst | dateCreated | seq | dataID | +--------------------------------+------+--------------------------------+-------------+-----+--------+ | PHID-FILE-qffup4ki3h7houulve3d | 26 | PHID-FILE-qffup4ki3h7houulve3d | 1430992455 | 1 | NULL | | PHID-FILE-qffup4ki3h7houulve3d | 25 | PHID-FILE-qffup4ki3h7houulve3d | 1430992455 | 0 | NULL | | PHID-FILE-qffup4ki3h7houulve3d | 21 | PHID-USER-j4owr5kdhj243qlkhu6l | 1430992319 | 0 | NULL | | PHID-FILE-qffup4ki3h7houulve3d | 26 | PHID-TASK-mirzruisn77tw4zow52a | 1430991972 | 0 | NULL | | PHID-FILE-7oogy4flcolzbmxzoqy7 | 26 | PHID-TASK-aktaetetss24d24lnzsu | 1430930202 | 0 | NULL | +--------------------------------+------+--------------------------------+-------------+-----+--------+ 5 rows in set (0.00 sec)
Yep, should be safe (although the first two rows are probably the issue?).
I should have a patch shortly unless this turns out to be harder than I think.
Deleting the rows resolved all the issues in T8119: Strange behaviour (possibly infinite recursion) on feeds and tasks.
Cool. D12756 should resolve this in the general case by terminating loops as soon as we try to re-load an object we're already loading.