mirror of
https://github.com/logos-messaging/logos-delivery.git
synced 2026-07-17 18:19:36 +00:00
77cb8a6c7a
* refresh merkle proof cache reactively on lightpush rejection
* Add tests for lightpush force refresh regenerates proof
* Fix linting
* Add nonce rollback for lightpush publish retry
* Fix nonce rollback race: make rollbackNonce CAS-conditional
Two bugs in the previous blind-decrement rollbackNonce():
1. Concurrent draw: another proof generated between attempt 1 and the retry
advances nextNonce, so decrementing steals a nonce already claimed by
another live message — nullifier collision, on-chain slashable.
2. Pre-attached proof: if the incoming message already carries a proof,
checkAndGenerateRLNProof short-circuits and draws no nonce, yet the retry
still called rollback, stomping on whatever nonce the preceding message
legitimately drew.
Fix: rollbackNonce now takes the specific nonce as a parameter and only
decrements when nextNonce == nonce + 1. checkAndGenerateRLNProof returns
(msg, messageId) so callers can thread the drawn id through to the retry
as reuseMessageId; generateRLNProof no longer performs an implicit rollback.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Extend reactive Merkle proof refresh to REST relay and broker paths
With PathCheckMinInterval removed, a non-empty merkleProofCache is trusted
forever unless force=true is passed. Previously only the two lightpush retry
paths ever passed force=true; the REST relay publish handlers (static and
auto-sharding) and the broker proof provider had no rejection feedback loop,
so a sliding root window would permanently reject their messages until restart.
- REST relay handlers: run validateMessage first; if it returns
RlnValidatorErrorMsg, force-refresh the cached path and retry validation once
before publishing — mirroring the lightpush reactive pattern.
- Broker provider (rln.nim): decode the generated proof bytes, call
validateRoot on the embedded Merkle root, and force-refresh + regenerate
when the root is outside the acceptance window — since the broker has no
external rejection signal to react to.
Tests added for all three new paths, using the corrupted-cache technique
(all-zero merkleProofCache → garbage root → rejection → retry).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Gate lightpush Merkle refresh on RlnValidatorErrorMsg for 420 responses
420 INVALID_MESSAGE is returned for any validateMessage failure, not just
RLN ones (e.g. oversized messages), so a bare status-code check triggers
an unbounded on-chain fetchMerkleProofElements RPC per rejected message.
Match the legacy-lightpush path: require the error description to contain
RlnValidatorErrorMsg before refreshing; 504 OUT_OF_RLN_PROOF remains
unconditional as it is unambiguously RLN-specific.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix merge conflict residue: replace RlnResult with Result in proof.nim
RlnResult[T] was removed as a type alias in commit 0a1700e23 (replace RLN
specific Result with generic Result). The merge left one stale reference at
proof.nim:62; replace with the equivalent Result[seq[byte], string].
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix ResultDefect on success and stale test stubs in lightpush retry path
Two bugs left by a bad merge:
1. lightpush.nim: isRlnRelatedFailure was computed before firstResult.isOk()
was checked, accessing firstResult.error unconditionally. When publish
succeeds, this raises a ResultDefect. Reorder the guard so isOk()/isNone()
short-circuits before the error fields are touched.
2. test_wakunode_lightpush.nim: the 420-retry stubs used descriptions
("simulated stale merkle path", "still stale") that don't contain
RlnValidatorErrorMsg. The code gates 420 retries on that substring to
distinguish RLN rejections from unrelated INVALID_MESSAGE responses
(e.g. oversized messages). Update stubs to emit RlnValidatorErrorMsg so
the retry path fires as the tests expect.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix linting
* Dedupe checkAndGenerateRLNProof force-refresh test
Same test existed in both test_wakunode_lightpush.nim and
test_wakunode_legacy_lightpush.nim. It exercises the checkAndGenerateRLNProof
primitive (short-circuit override on force=true + nonce rollback +
cache refetch), not lightpush-specific behavior. The modern lightpush copy
provides full coverage; the legacy copy paid the anvil/on-chain setup cost
for zero incremental coverage.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Bound lightpush RLN refresh-retry with a 5s timeout
The reactive refresh + retry path calls into the RLN contract (eth_call for
getMerkleProof) then republishes over libp2p — neither has an outer bound,
so a hanging RPC endpoint could stall the caller indefinitely. Wrap the
retry in withTimeout(RlnRefreshRetryTimeout); on timeout, log and return
the original error rather than hang. Applied to both the legacy and modern
lightpushPublish retry paths.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Extract broker RLN proof provider into generateRLNProofWithRootRefresh
The broker proof provider's inline logic — generate proof, decode it,
self-validate the root against the acceptable-root window, and on stale
root force-refresh + regenerate — is a reusable primitive on its own.
Lift it into rln/proof.nim next to the existing generateRLNProof so the
provider closure in mount() collapses to a single call, and any future
non-broker caller can share the same flow.
Pure refactor; behavior identical.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Extract REST relay handlers' RLN attach+validate+retry into a helper
Both the static and auto POST handlers duplicated the same block: attach an
RLN proof, run wakuRelay.validateMessage, and on RlnValidatorErrorMsg
substring force-refresh the merkle path and revalidate. Lift the block into
attachRlnProofValidateWithRetry alongside a small RlnPublishError variant
so the handlers can keep dispatching 500 for proof-gen failures and 400
for validator rejections. Each handler is now ~15 lines instead of ~35 and
the retry contract lives in one place.
Pure refactor; behavior unchanged. Verified against
tests/wakunode_rest/test_rest_relay.nim (13/13) including the two stale-
RLN-proof retry tests added earlier on this branch.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Simplify legacy lightpush autosharding path to a direct getShard call
The single-content-topic path went through getShardsFromContentTopics,
iterated the resulting table under the assumption it was 1-entry, and
had a defensive fallthrough (`resolved = false` → bare `return`) for
the empty-table case. That case is unreachable: on success the table
always has exactly one entry, and the fallthrough returned a default-
initialized Result — not behavior worth preserving.
Replace with the same two-step pattern the modern lightpushPublish uses
(NsContentTopic.parse → sharding.getShard) so both handlers share a
shape and the (now-explicit) failure paths surface useful error strings.
Verified against tests/node/test_wakunode_legacy_lightpush.nim (9/9).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Break up legacyLightpushPublish into single-purpose helpers
legacyLightpushPublish had grown to ~100 lines with three concerns
interleaved: choosing between client/self-request publish, resolving the
effective pubsub topic, and running the RLN refresh-retry with timeout.
Lift each into a private top-level proc so the main proc reads top-to-
bottom in ~35 lines: precondition → prepare message + proof → resolve
topic → publish once → RLN-refresh retry if applicable.
The new helpers are:
- internalLegacyLightpushPublish — client vs self-request dispatch
(was the inline closure)
- resolveLegacyPubsubTopic — explicit param or autosharding
- runRlnRefreshRetry — force-refresh proof + retry with
RlnRefreshRetryTimeout bound; returns caller's fallback on timeout
Nonce now appears in a top-level signature, so import it directly from
rln/nonce_manager rather than re-exporting through rln/proof.nim (the
latter would leak nonce_manager's bulk chronos/times exports and clash
with waku_rendezvous overload resolution).
Pure refactor; behavior unchanged. Verified against
tests/node/test_wakunode_legacy_lightpush.nim (9/9) and
tests/node/test_wakunode_lightpush.nim (11/11).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Replace forceMerkleProofRefresh flag with invalidateMerkleProofCache
The forceMerkleProofRefresh bool was threaded through five procs
(checkAndGenerateRLNProof → generateRLNProof → generateProof base + on-chain
→ ensureFreshMerkleProofPath) purely so a caller could ask the merkle path
to be refetched from chain on a stale-RLN retry. Reading any one of those
signatures required a mental stack trace back up to the retry caller to
know what "force" meant.
Flip the contract: expose invalidateMerkleProofCache on the group manager;
callers invoke it on the retry path to empty the cache, and the normal
proof-gen flow then sees an empty cache and refetches on its own. No flag
crosses procs.
Removed forceMerkleProofRefresh from:
- group_manager_base.generateProof (+ on_chain override)
- ensureFreshMerkleProofPath (its `force` param)
- generateRLNProof
Renamed on checkAndGenerateRLNProof: forceMerkleProofRefresh → regenerate.
It now controls one thing — bypassing the "message already has a proof"
short-circuit — and never propagates downstream. Callers pair it with
invalidateMerkleProofCache when they need the refetch semantics too.
Retry callers updated in lockstep: generateRLNProofWithRootRefresh, REST
attachRlnProofValidateWithRetry, and both legacy + modern lightpush retry
paths now call groupManager.invalidateMerkleProofCache() before regenerating.
The primitive test in test_wakunode_lightpush.nim now exercises the pair
directly, and the onchain ensureFreshMerkleProofPath test is rewritten to
exercise invalidate → ensureFresh as one flow (retiring its `force = true`
variant).
Pure refactor; behavior unchanged. Verified against:
tests/waku_rln_relay/test_rln_group_manager_onchain.nim (29/29)
tests/wakunode_rest/test_rest_relay.nim (13/13)
tests/waku_rln_relay/test_wakunode_rln_relay.nim (5/5, 1 skip)
tests/node/test_wakunode_lightpush.nim (11/11)
tests/node/test_wakunode_legacy_lightpush.nim (9/9)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Reject empty merkle proof cache in on-chain generateProof
An empty cache passes the existing mod-32 length check and silently
produces a witness with no path elements, yielding a proof no validator
will accept. Catch the case explicitly so the failure surfaces at the
callsite rather than after a wasted publish round-trip.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Drop nonce reuse on RLN retry; consume a fresh nonce instead
The retry path threaded a rejected attempt's message id back so
NonceManager.rollbackNonce could CAS the counter and reclaim the slot.
This bought at most one nonce per retry (a rare event) at the cost of
an Option[Nonce] param, a tuple return, and a concurrent-draw race the
CAS had to defend against.
Simpler to optimistically waste the nonce: checkAndGenerateRLNProof
returns Result[WakuMessage, string] and callers drop the tuple
destructuring. NonceManager.rollbackNonce and its tests, now dead,
are removed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Split checkAndGenerateRLNProof: extract attachRLNProof
The `regenerate: bool` param was control coupling: every caller passing
`regenerate = true` was a retry path that had already established
`rln.isSome()` (it just called `invalidateMerkleProofCache` on the
group manager) and wanted an unconditional fresh proof, not a "check
and maybe generate".
Extract that into `attachRLNProof(r: Rln, message)`, which always draws
a fresh nonce and rebuilds the proof, replacing any existing one.
`checkAndGenerateRLNProof(rln: Option[Rln], message)` keeps only its
"check" behaviour — short-circuit on an existing proof, pass through
when RLN is not configured — and delegates to `attachRLNProof`.
Taking a concrete `Rln` deletes the previously unreachable
`regenerate = true` + `rln.isNone()` branch (which would have
republished the stale rejected proof) at the type level, and removes
the duplicated nonce-draw/generateProof/encode body that
checkAndGenerateRLNProof carried alongside generateRLNProof.
The two lightpush retry sites and the invalidate+regenerate test now
call attachRLNProof directly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Bound only the proof refresh on lightpush RLN retry, not the republish
RlnRefreshRetryTimeout previously wrapped the whole retry leg: merkle
path refetch + proof regeneration + the second publish. Sharing one
clock meant a slow (but successful) refetch could leave no budget for
the publish, cancelling it mid-flight — burning the freshly drawn nonce
and, worse, reporting the original RLN rejection while the retried
request may already have reached the service node and been relayed.
Move the withTimeout to cover only attachRLNProof (the on-chain
eth_call + proof regeneration — the one step with an unbounded external
dependency). The retried publish runs unbounded after it, matching the
first attempt's semantics, so slow transports (e.g. an optional mix
route) cannot be cancelled by the refresh budget and the caller always
receives the publish's actual result.
On refresh timeout the retry publish never starts and the original
rejection is returned, so a "failed" answer now guarantees no second
request is in flight.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Shield the shared merkle path refetch from caller cancellation
ensureFreshMerkleProofPath coalesces concurrent callers onto a single
in-flight refetch future and awaited it directly. Chronos propagates
cancellation into the awaited future, so one caller being cancelled
(e.g. a publish retry whose RlnRefreshRetryTimeout fired mid-refetch)
cancelled the shared future under every coalesced caller — and threw
away an eth_call that may have been about to complete.
Await the in-flight future via join() instead: cancelling a caller only
detaches its callback. The refetch always runs to completion, coalesced
callers keep waiting, and even when every caller times out the cache
still gets populated in the background so the next publish fast-paths
instead of paying the RPC round-trip again.
Detaching is safe because doRefresh cannot fail with an unhandled
exception: the eth_call wrappers catch CatchableError and return
Result, and failures surface to callers through fetchOk / an empty
cache exactly as before.
Adds a test cancelling the initiating caller mid-refetch and asserting
the shared future survives and the coalesced caller completes with a
populated cache.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Move message publish retries out of the kernel on stale RLN proof
On a stale-merkle-proof rejection the kernel and REST layers no longer
republish internally. They schedule a detached merkle proof refresh and
fail fast with a distinguishable, retry-worthy error, leaving the retry
decision to the caller.
- add GroupManager.scheduleMerkleProofRefresh: drops the cached path and
asyncSpawns the refetch (join-shielded, so a cancelled caller never
cancels the shared refetch); base is a no-op.
- v3 lightpush (node.lightpushPublish): on an RLN-related 420/504, call
scheduleMerkleProofRefresh and return immediately, normalized to 504
OUT_OF_RLN_PROOF with RlnProofRefreshScheduledMsg. The internal
republish and its bounded timeout are gone.
- send service: lightpushPublishToAny now routes through
node.lightpushPublish so every NextRoundRetry round regenerates the
proof against the current cache; the existing OUT_OF_RLN_PROOF ->
NextRoundRetry mapping is the retry.
- REST lightpush: pass the 504 through unchanged (no internal retry).
- REST relay: attachRlnProofValidateWithRetry -> attachRlnProofAndValidate;
validate once, on RLN-invalid schedule the refresh and return 503 with
the marker instead of regenerating and revalidating in-request.
- rename RlnRefreshRetryTimeout -> RlnMerkleProofRefreshTimeout; it now
bounds only the legacy lightpush refresh, which keeps its one-shot
internal retry (no wire error taxonomy to delegate to callers).
Not touched: legacy lightpush retry behavior, the proof.nim
attachRLNProof/checkAndGenerateRLNProof split, and the merkle path
refetch/coalescing internals of ensureFreshMerkleProofPath.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Tighten RLN proof-refresh docstrings and comments
Condense the verbose comments introduced with the reactive merkle
proof-refresh work. No code changes.
- group_manager_base: base methods carry the contract; the
OnchainGroupManager overrides keep only their concrete behavior
instead of repeating the rationale.
- proof.nim, lightpush.nim, constants.nim: trim multi-line docstrings
and inline comments to their load-bearing facts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Make the merkle proof path refresh safe against concurrent invalidation
ensureFreshMerkleProofPath returns the fetched path instead of leaving
callers to re-read merkleProofCache. generateProof uses the returned
value, so an invalidate landing during the refresh's updateMemberCache
suspension can no longer empty the cache under an in-flight proof-gen
and trip its empty-cache guard. The cache field is now only a hint for
the next call.
invalidateMerkleProofCache also bumps merkleProofCacheGeneration. The
refresh stamps the generation before each fetchMerkleProofElements and
only caches a result whose generation is unchanged, otherwise it fetches
again, bounded by MerkleProofRefetchMaxAttempts. A rejection-driven
invalidate is therefore never absorbed by a fetch that predates it, so
the path a caller receives always postdates the invalidate.
refreshRoots awaits rootsRefreshInFlightFut via join() at both await
sites, giving the roots refresh the same cancellation shield the proof
path refetch already had: a cancelled validateRoot caller no longer
cancels the refresh its coalesced peers are waiting on.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>