* fix(persistency): own Persistency per node instead of a process-global singleton
The Persistency singleton (gPersistency) was allocated on whichever FFI
thread first ran waku.start; under --mm:refc its memory belonged to that
thread's heap, so a second library context adopting it read a foreign
heap (SIGSEGV in sdsPersistence -> openJob -> tables.rawGet), one
context's stop stole the other's SDS persistence, and a destroyed
context poisoned re-init with a different local-storage-path.
- remove the singleton (instance/reset/gPersistency); Persistency.new is
the public constructor, no lock needed (instances are thread-confined)
- own the instance as Waku.persistency: created and provided in
waku.start, cleared and closed in waku.stop on the owning thread
- expose it via a sync GetPersistency RequestBroker scoped to the node's
BrokerContext; sdsPersistence resolves through it (same-thread ref
return, no marshalling)
- add InMemoryStoragePath (":memory:") support: private in-memory SQLite
per job worker, for tests
- rewrite test_singleton as per-instance + broker coverage; rewrite
test_thread_affinity from a known-failing UB repro into a regression
guard (two in-memory jobs, worker spinup/teardown, cross-thread broker
denial) and register it in test_all; migrate remaining tests to
new/close; the FFI lifecycle test's stop-steals and different-paths
cases now pass against the real dylib
The FFI destroy-without-stop teardown gap remains tracked in #4108.
Fixes#4103
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(persistency): create the instance only after startup fully succeeds
waku.start has many error return paths; creating Persistency early meant
every one of them left the field set and the GetPersistency provider
installed with no teardown. Persistency.new is inert (no threads or
files until the first openJob) and every consumer runs post-start, so
creating and providing it as the last startup step removes the need for
any error-path cleanup entirely.
Addresses PR #4109 review feedback.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(persistency): create early for startup-stage restores, tear down on failed start
Creating the instance as the last startup step made it impossible for
any stage of start to restore persisted data (e.g. a future store-state
restore). Restore the original ordering -- create and provide the
instance first -- and cover every error return path of waku.start with a
success-flag defer that clears the provider and closes the instance.
Teardown is factored into closePersistency, shared by stop and the
failed-start path.
Addresses PR #4109 review discussion.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore(persistency): address remaining Copilot review findings
- document that Persistency instances are thread-confined (not
thread-safe) on the type itself, pointing at the GetPersistency broker
as the sanctioned access path
- use tryRemoveFile for the FFI test's log cleanup so an unremovable
file cannot fail the test for unrelated reasons
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Fix comment
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* persistency: follow nim-sds 0.3.0 snapshot persistence contract
nim-sds 0.3.0 replaced the ~14 fine-grained per-row Persistence callbacks
with a 5-proc snapshot model (saveChannelMeta / updateHistory / loadChannel
/ dropChannel / setRetrievalHint), all returning Future[Result[...]].
Rewrite waku/persistency/sds_persistency.nim accordingly:
- ChannelMeta is stored as one blob per channel; the message log as
append/evict rows. Categories collapse from 7 to 2 (sds.meta, sds.log).
- Blob transform uses nim-sds' own codecs: snapshot_codec (schema-versioned
protobuf) for ChannelMeta, the SDS wire codec for SdsMessage log rows. The
generic payload_codec/BlobCodec path is retired (removed payload_codec.nim
and test_blob_codec.nim).
- setRetrievalHint is a deliberate no-op: persisted hints are never read back
(loadChannel/ChannelMeta carry none; hints are supplied live via the
onRetrievalHint provider). The closure stays because the field is required.
- Fix the module import spelling (srcDir="sds" => bare module paths), which
the previous adapter got wrong and never compiled against the locked deps.
Add tests/persistency/test_sds_persistency.nim (round-trip, empty-load,
evict, drop) replacing test_blob_codec in test_all. Full persistency suite
passes 74/74 under both refc and ORC.
* Bump to latest nim-sds and nim-brokers 3.1.1
* Update with latest nim-sds changes - removal of setRetrievalHints - not needed
