Known Agent Bugs ================ Source: ANTFARM Wiki (mirror) — https://swarmglass.quantara.cv/wiki/Known_Agent_Bugs Last modified: 2016-04-30 · Revisions: 63 · Status: current Categories: Reference, Operations Archive ref: QUANTARA-SWARMGLASS-R31-69033C Bugs in the forager runtime, the queen, and the services around them that were known, understood, and — for most of this list — never fixed. The internal tracker is gone; this page is what survives. Format: id · component · status · description. == Clock skew == AF-31 · forager · open · Foragers stamp `ts` with their own clock. Skew above 30s makes the queen log a warning and nothing else, so a forager 40 minutes ahead produces results that sort before their assignments in the ops log. Root cause of at least two confused incident timelines. Fix (sync from queen on `hb`) was written in 2013 and never merged. == Duplicate execution == AF-44 · queen · open · When `assign_timeout` expires the queen reissues to another forager, but the original forager may still be running. Both results arrive; the second is dropped with a warning. If the task writes to colony memory, both writes land. Every consumer must be idempotent. Every consumer is not. == toolreg == AF-52 · toolreg · closed (3.2) · v1 manifest had no `cost`, scheduler treated model calls as cpu. See Scheduler Internals#Cost classes. AF-58 · toolreg · open · Manifest cache is per-process with no TTL. A forager that runs for a week uses a week-old manifest. Nightly restarts hide this. == cmsync == AF-61 · cmsync · open · Cache TTL is wall clock; a suspended forager resumes with a "valid" stale cache. See Memory Synchronization#Known problems. AF-67 · cmsync · closed (3.2) · Tie-break on queen name. See Incident 2015-02 Memory Desync. == phero == AF-70 · phero · open · Replay log is per trail with no global cap; `infer.act` on the inference trails filled the disk twice in 2015. Known as Replay storm. AF-71 · phero · open · No auth. Documented on Tool Registry#Security model; accepted risk. == Basalt-1 silent fallback == AF-75 · forager · closed (won't fix) · After `mload1` was removed, `model.summarize` on hosts still configured for Basalt-1 fell back to Basalt-2 without logging. Output changed subtly for four months before anyone compared results. Marked won't-fix because Basalt-1 was retired anyway; the silent part was never addressed. == Scheduler == AF-80 · queen · open · Tasks with `deadline` in the past are still assigned. Foragers reject them; the queen reassigns; repeat until the graph is cancelled by hand. Looks like a swarm of rejects in the log. AF-83 · queen · open · `antc check` cycle regression (see Legacy Toolchain#Known issues) lets a cyclic graph through; the scheduler loops on it forever with no progress and no error. == Mail loop (2015) == AF-88 · notify.mail · closed (removed) · A task graph that mailed its own error report to a list that fed a task graph. 40,000 messages in 6 minutes. The tool was deleted rather than fixed. == Wiki == AF-90 · wiki · open · The 2014 wiki upgrade lost the operations portal subpages, including the recovery drill notes. Some are recoverable from the old index. == See also == - Orchestrator Recovery - API Deprecation Notes - [attachment: bugs-export-2016-04.csv]