Fix notify flood memory growth #611

Merged

paratoxic merged 3 commits from jobs-dos into master

2026-08-23 16:13:59 +00:00

paratoxic commented

2026-08-22 20:02:58 +00:00

Owner

Fixes this:

M4. Proxy: notify flood memory growth — `src/jobs.rs:35-48`

ckpool: ages out workbases older than 600 s, keeps at least 3, skips
in-use (stratifier.c:1081-1093). No rate limit.

Design:

Cap Jobs.valid at a constant (64 jobs). On insert, evict the oldest
by JobId order when the cap is exceeded. Evicted job -> share gets
StratumError::Stale (already the unknown-job path at
stratifier.rs:766; does not touch the bouncer).
Count cap vs. ckpool's time window is a deliberate divergence: a time
floor reintroduces unboundedness under flood; count is the correct
simple bound and serves as backpressure.
Identical accumulation shape in pool mode (height-based clean means
templates pile up per --update-interval within a block gap); the cap
fixes both.
Local monotonic JobIds (upstream IDs stored as upstream_job_id in
Job) make eviction order = insertion order; latest (max ID) is
never evicted, so assert_invariants holds by construction.
Must respect pinned tests: multiple_jobs_accumulation (5 jobs
retained) and assert_invariants (latest in valid). Cap >= 5
keeps all existing tests green; 64 is generous (~10.7 min of history
at default 10 s update interval, near ckpool's 600 s window).
Keep seen unbounded per seen_never_evicts_within_job_group —
growth is PoW-gated at 32 bytes per accepted share.
Implemented as BTreeMap + pop_first eviction (precedent:
orphan_receipts: BTreeMap, router.rs:53); unit test
check_evicts_oldest_beyond_cap covers both workbase types.

Fixes this: ## M4. Proxy: notify flood memory growth — `src/jobs.rs:35-48` **ckpool:** ages out workbases older than 600 s, keeps at least 3, skips in-use (`stratifier.c:1081-1093`). No rate limit. **Design:** - Cap `Jobs.valid` at a constant (64 jobs). On insert, evict the oldest by `JobId` order when the cap is exceeded. Evicted job -> share gets `StratumError::Stale` (already the unknown-job path at `stratifier.rs:766`; does not touch the bouncer). - Count cap vs. ckpool's time window is a deliberate divergence: a time floor reintroduces unboundedness under flood; count is the correct simple bound and serves as backpressure. - Identical accumulation shape in pool mode (height-based clean means templates pile up per `--update-interval` within a block gap); the cap fixes both. - Local monotonic `JobId`s (upstream IDs stored as `upstream_job_id` in `Job`) make eviction order = insertion order; `latest` (max ID) is never evicted, so `assert_invariants` holds by construction. - Must respect pinned tests: `multiple_jobs_accumulation` (5 jobs retained) and `assert_invariants` (`latest` in `valid`). Cap >= 5 keeps all existing tests green; 64 is generous (~10.7 min of history at default 10 s update interval, near ckpool's 600 s window). - Keep `seen` unbounded per `seen_never_evicts_within_job_group` — growth is PoW-gated at 32 bytes per accepted share. - Implemented as `BTreeMap` + `pop_first` eviction (precedent: `orphan_receipts: BTreeMap`, `router.rs:53`); unit test `check_evicts_oldest_beyond_cap` covers both workbase types.

paratoxic added 1 commit

2026-08-22 20:02:59 +00:00

Fix notify flood memory growth

CI / ckpool (pull_request) Successful in 27s

Details

CI / lint (pull_request) Failing after 3m5s

Details

CI / test-linux (pull_request) Failing after 6m40s

Details

79d11de081

paratoxic requested review from parabit

2026-08-22 20:03:03 +00:00