Changes touching this path
- the error shipper (ADR 0101): every minute it reads the wide-event lines out of each named container Docker json-file log and writes them into loot_metrics, migration 0004: wide_event for 30 days, error_issue with no identifier, wide_counter where a minute printed twice adds, and wide_log_position, under a new errors_writer role, since the beacon writer role may never read an issue. issue.ts fingerprints by source, type and the top three in-app frames and regresses a resolved issue on an occurrence from another release after the resolve; db.ts applies events, issues, counts and the position in one transaction, so a batch read twice counts once; alert.ts posts to Discord at most once an hour per issue with the type, culprit, route, release and count and never the message; shipper-main.ts finds containers by the name in config.v2.json and fails when it finds none. bundled to .output/errors.cjs; ci/test-main.sh provisions the role. pinned in errors-ship.test.ts, 15 tests, and errors.pg.test.ts, 5 over Postgres 18; red under named mutations, each restored, failed and passed: a regression on the same release (1 and 13, and 1 and 8 on pg), the route in the fingerprint (1 and 13), vendor frames fingerprinted (1 and 13), the title in an alert (1 and 13), no hourly limit (1 and 13, and 1 and 8), a partial entry kept (1 and 13), consumed past the last line (1 and 13), a re-read counted again (1 and 8), a minute replaced not added (1 and 8), the beacon reading issues (1 and 8), and no expiry (1 and 8, after the test counted the rows left, where it first passed). an end-to-end run of the built bundle over the built server output made two issues and two Discord posts with no message. site gate green at 1063 passed. owes a site deploy with the scripts change (vuxxwrys)
9fdc1ae3 · dbf3dbe6… - a leaked secret in stored error text can be purged (ADR 0101): loot_metrics migration 0005 defines purge_error_text, callable only by the migrating superuser, which deletes every wide_event row holding the text as a literal substring in any text column, cuts a matching error issue title back to its type and keeps its fingerprint and counts, answers two counts, and refuses a text under 8 characters; the scripts repo errors:purge verb takes the text masked or from a file, never a command line. it also joins the lines Docker json-file splits past 16 KiB, which the first production tick of the error shipper dropped as partial=1 and which for a busy minute counter line loses the exact counts: parseDockerLog walks the bytes, joins consecutive entries per stream, drops a joined line only past 1 MiB, and reports safeBytes so the position never passes a line still being split. pinned in errors.pg.test.ts, 7 tests on Postgres 18, and errors-ship.test.ts, 17; red under named mutations, each restored, failed and passed: no delete, the title kept, anyone may purge, no length floor, the error text not searched (each 1 and 6), no join (1 and 16), the position past a split line (1 and 16, after the test put a whole stderr entry behind the split one, where it first survived), one pending line for both streams (2 and 15). owes a site deploy with the scripts change (ntkmpoqn)
568ee1cd · dbf3dbe6…diff - the error shipper reads every container labelled build.loot.wide-events, where it read a list of names, and reads at most 8 MiB of one log a run, resuming from its position and widening a window that holds no whole line, so the forge log, never rotated, is read down over several runs; a run that finds nothing new records the check on the position, so wide_log_position says when the shipper last ran. the scripts half creates the site and forge containers with the label and a rotated log, and the forge with LOOT_FORGE_RELEASE. pinned: the bound, the widening, the check, the label filter; red under named mutations, each restored. site gate 1096 passed, errors pg suite 15 passed over Postgres 18 (mwvpnyss)
Perf-Baseline: reset object_gets and graph_sorts moved with 587f93b, whose negotiation workload now builds the forge push a land makes (opywoozo); this change touches no Rust
4cf1d1e1 · dbf3dbe6…diff - an open error issue whose last hour holds ten or more occurrences and five times its hourly rate over the day before is alerted as spiking, at most once an hour, marked in the statement that finds it; the first shipper run from 14:00 UTC claims the day in error_digest (migration 0009) and sends the day's digest, the counts beside the day before, new and open issues and the three busiest by type and culprit, never a message. pinned over Postgres 18 and the alert text, red under named mutations, each restored (stzwopuy)
Perf-Baseline: reset object_gets and graph_sorts moved with 587f93b, whose negotiation workload now builds the forge push a land makes (opywoozo); this change touches no Rust
ef41eb8c · dbf3dbe6…diff
Renames are not followed. loot's tree maps a path to an address, so a rename is a delete and an add. This list is the history of the name, not of the bytes.