feat(webapp,run-engine,run-store,redis): wire the execution-snapshot store behind an off-by-default dial - #4783
Open
d-cs wants to merge 128 commits into
Open
feat(webapp,run-engine,run-store,redis): wire the execution-snapshot store behind an off-by-default dial#4783d-cs wants to merge 128 commits into
d-cs wants to merge 128 commits into
Conversation
The decorator that dual-writes snapshots to Redis has to own the snapshot id, or the same snapshot carries a different id in each store and the comparator chases a difference that is not real. Four of the six snapshot input types had no id field, so four write sites could not carry one. Add it to CompletionSnapshotInput, ExpireSnapshotInput, RescheduleSnapshotInput and CreateExecutionSnapshotInput, and thread it through every nested create. createCancelledRun built its create inline and dropped the id its input already carried; it now passes it too. The field is optional everywhere, so an absent id still falls through to Prisma's @default(cuid()) and no existing caller changes.
…ators RunStore has 71 members. A decorator that intercepts a dozen of them should not restate the other 59 forwarders alongside its real logic, and hand-writing them invites a typo no test would catch. Generate the base from the interface instead. The generator also emits the member-name lists, so the suite can assert that the class and the interface hold exactly the same members: a method added to RunStore and not to the base fails a test rather than becoming a silent hole in the decorator. The one data property on the interface becomes a getter over the delegate, read live rather than captured, so a delegate whose client changes is not cached.
…parity tests No nested write site returns the snapshot it created: createRun returns the run, expireParkedRun returns a count, and the rest return a selected TaskRun. So the Redis entry is built from each site's own input plus the caller-minted id. That means every value Postgres derives rather than receives has to be reproduced: the DEQUEUED-to-PENDING rewrite, the four values lockRunToWorker hard-codes, the three rescheduleRun defaults, and the engine column default a completion leaves unset. The parity suite covers all ten physical write sites, comparing the built entry against the row Postgres actually wrote. It caught the dropped id in createCancelledRun.
…write The last dial position makes the Redis store the sole snapshot writer, so Postgres has to stop writing snapshot rows without changing anything else it does. One constructor flag does that across all ten write sites. With it off, the nine nested creates are omitted and the run mutation still lands; createExecutionSnapshot echoes its input in the shape callers expect rather than inserting; and the completed-waitpoint join inserts are skipped, since they would otherwise link to a row that no longer exists. Defaults to true, so every existing caller and test is unaffected.
A decorator over any RunStore that also writes execution snapshots to Redis. It overrides only the methods that touch a snapshot and inherits the rest. Write order is the correctness property, and the two orders differ on purpose. A transition writes Postgres first: a crash in the gap leaves a stale latest snapshot, which the heartbeat stall watchdog already heals. A birth writes Redis first: a crash there leaves an unreachable key for a run that does not exist, where Postgres-first would leave a run with no snapshot at all and no way to read one. Each order is chosen so the crash state is the harmless one. A failed transition append retries three times, then hands the run to the repair job. It never rethrows, because Postgres has already committed and a throw would turn a healable gap into a caller-visible error. A failed birth append is survivable before redis-only, where Postgres still holds the snapshot, and refuses at redis-only, where it would otherwise create a run with no snapshot anywhere; refusing works only because the birth append comes first. None of the four non-failure append outcomes enqueues a repair: an absent keyspace is every pre-cutover run's transitions, a fork means another writer advanced the head, a duplicate is a retry that landed, and a cycle mismatch is the store refusing an untrustworthy pointer on purpose. At mode off the decorator makes no Redis call and builds no entry.
…ore handles Proves the deferral from inside the transaction callback rather than assuming it: a staged append is absent from Redis while the transaction is open and present once it commits, and a rollback leaves both stores agreeing the transition never happened.
The engine resolves its since-cursor to a createdAt before it asks for the window, so the snapshot id is gone by then and the id-addressed read cannot serve it. Adding a cursor-addressed read is the alternative to changing the engine's read path, which stays untouched. The cursor is exclusive and keeps the same-millisecond blind spot the Postgres read has. Matching it is the requirement, not an oversight: a Redis read that is more correct than the Postgres read shows up as divergence during compare mode, which exists to surface real defects. Closing the blind spot needs seq ordering on both sides and belongs after the cutover. The walk goes newest-first and stops at the first entry at or before the cursor, so its length is the length of the answer rather than the run's history. This adds a read operation. It does not touch the append script, the keyspace, or the write-ordering protocol.
…back Two of the five snapshot reads take arbitrary Prisma arguments, and a key-value store cannot answer an arbitrary query. Only three production call sites exist, all in the engine's executionSnapshotSystem, and both generic ones send a single fixed shape, so the decorator recognises exactly those shapes and delegates everything else. Each matcher rejects an argument object carrying a key it does not know, because a query that has drifted must be answered correctly by Postgres rather than approximately from Redis. A miss is the coexistence path, not an error: a pre-cutover run or expired history falls back to Postgres. The entry supplies every scalar column, and the checkpoint and waitpoint rows are read back through the delegate only when the entry says they exist, so the common read of a running run makes no Postgres call at all. Which runs read from Redis is a hash of the run id, so a run does not change store between two reads of one poll, two instances of the same dial agree, and raising the dial only ever adds runs to the cohort.
Two rules, because neither can see what the other leaves behind. A terminal run whose keyspace never received the completion expiry gets one applied, so it reaps on the schedule a healthy terminal append would have set. A keyspace with no run row at all, past an age threshold, is deleted outright — that is a crashed birth, which is non-terminal so it carries no expiry and has no run row, so the first rule can never match it. It never reaps on an unknown answer: a live run is left alone however old its keyspace, a young orphan is left for the birth that may still be in flight, and a batch whose run lookup failed is skipped rather than treated as absent. Run rows are resolved through the run store rather than a raw client, because under the run-ops split a run can live on either database and a raw lookup would report a live run as an orphan. Nothing schedules this. The engine's worker has to run it, and run-store cannot reach the engine. Also moves the decorator suites onto the worker-scoped container fixture. The per-test one boots a Postgres and a Redis container for every test, which is what the replication tests need and these do not; the sweeper suite alone went from repeated two-minute timeouts to ten seconds.
…eads on The engine's own flows, driven against the decorator with every snapshot read served from Redis, injected through the store seam that runStoreInjectability already proves. Same flows, same expectations, different store underneath — the point is that nothing in the engine has to know, so no existing suite changes. Covers a run driven to completion, the execution data at each step, a since-window wider than the fifty cap, and a pre-cutover run with no keyspace falling back to Postgres. The environment-boundary test asserts parity rather than a fixed shape: whatever Postgres answers for a foreign environment, Redis has to answer the same, or the tenant boundary behaves differently once reads move over.
…oth stores Three defects, all of which passed the existing suites because no test drove a snapshot that actually carried waitpoints, and because the parity suite compared createdAt against a value it had just read back from the row. The decorator never passed a cycle to the append, so no wp:<cycleSeq> key was written for any snapshot and the completed-waitpoint side of Redis was permanently empty. It now mints a cycle when the id set differs from the current head and carries the previous cycleSeq forward when it does not, so a resume writes the record set once and the copy-forwards that follow write no key at all. The since-window hydration returned an empty completedWaitpointOrder. That column is not the join: the engine reads it off the head row as the oracle that gives each completed waitpoint its position in a batch, so an empty order resumed every batched triggerAndWait with an undefined index. Seven of the eight write sites stamped the entry from the app clock while Postgres stamped its own column default, so the two stores held different instants for one snapshot. The decorator now supplies createdAt, and an equal updatedAt, at every site, and the standalone path supplies it too rather than reading the row back. Beyond making the field comparable, this aligns the since-window: the cursor is resolved from one store and applied in the other, and two different instants misfilter that window. The parity suite gains an independent clock-provenance guard, and a case proving an absent instant still takes the database default, which is what keeps the store's behaviour unchanged while the decorator is off.
…oint The generator that emits the pass-through store base is a runnable script, not dead code, and the same glob covers any script added there later.
…arity real The sweep discovered keyspaces by their cur key, which the append script writes only when an entry is valid. A keyspace whose entries all carry an error has no cur and no index, so neither sweep rule could ever see it and it leaked with no expiry, which is the same unbounded leak the second rule exists to close. It now scans on the entry hash, which every append writes, and the age probe falls back to the newest instant in that hash when the index is empty. Enumerating a run's cycle keys used KEYS. That command iterates the whole database and blocks while it does, and a hash tag routes a key without scoping the scan, so a sweep pass would have issued one full keyspace scan per run. It now reads the dense cycle high-water counter the append script maintains, which is the same source the store's own terminal-expiry loop uses, and pipelines the existence checks into one round trip. The timestamp parity assertion was still tautological. The previous commit added a note saying the builders receive an independent instant and did not change the builder calls, which kept reading the value off the row under test. Every case now mints one instant, passes it to the store, and gives the builder the same value, so a write site that stops forwarding the caller's instant fails here. Also documents what an injected fault actually does at each write path, since only the birth path rethrows, and scopes a run count in the chaos suite to the environment under test.
Review asked why the pass-through base is tested against a hand-built delegate rather than a real store. Checking what the compiler already guarantees showed the test's own stated reason was wrong, and that one of its cases could not fail. implements RunStore already rejects a missing member with TS2420, so the claim that a method added later would become a silent hole was not true. The case comparing the class against the generated name list could not detect a parse miss either, because both the class and the list come from one parse of the interface, so a miss drops the member from both sides. The generator's comment asserting otherwise was false. Parity now lives where it can actually fail: assertions tying the name lists to keyof RunStore in both directions, and one rejecting a public member the class declares and the interface does not. They sit in src rather than in a test, because the build config excludes test files, so a type assertion written in a test is never checked. Each was verified by making it fail. What the compiler cannot see is inside the forwarder bodies, since every one is typed (...args: any[]): any. A forwarder wired to the wrong member, or dropping an argument, typechecks cleanly. The remaining probe covers exactly that, using a per-member sentinel so a misrouted body returns the wrong value rather than merely returning something. Verified by rewiring a forwarder: typecheck passes, the probe fails and names the member. Renames the double to forwardingProbe across both suites and says at the top why a container cannot replace it: no database is involved in whether a pass-through passes through.
A member was removed from the generated list while verifying that the new parity assertion fails when one goes missing, and the restore did not run, so the verification state was committed. Regenerated from the interface. The assertion did its job: typecheck rejects the list, naming the missing member.
…ecorator-tri-13449
… client key prefix Two defects from review, both silent. An append staged inside a transaction dropped its expected-head argument, and the post-commit flush passed undefined in its place. That disabled the compare-and-set for every snapshot written inside a transaction, which is the path both engine transaction writers use, so a stale append that the store would have refused as forked was written instead and became the head. The expectation now travels with the staged entry. The sweep built its scan pattern without the client key prefix. ioredis prepends that prefix to keys for ordinary commands but not to a SCAN MATCH pattern, and returns matched keys with it still attached, so a prefixed client made the sweep match nothing and report a clean pass. The engine sets a prefix on every other Redis client it builds, so this would have surfaced at wiring time as a reaper that silently protected nothing. Also removes a keyPrefix option on the sweep that could never work: the keyspace prefix belongs to snapshotKeys in the store, which writes snap: keys unconditionally, so there was no other keyspace to point it at. Both fixes have a test verified by reintroducing the defect: the staged stale append is written without the guard, and the prefixed sweep scans nothing.
… source The generator was scaffolding for a one-off job: writing 70 near-identical forwarders. Keeping it meant carrying a hand-rolled scanner over the interface body, because the TypeScript compiler API is not resolvable in this workspace, which is more machinery than a file that changes only when the interface does. The two files it produced are now maintained by hand, and their headers say so. Nothing is lost, because the generator was never what guaranteed they were right. That is the compiler: implements RunStore rejects a missing member, and the parity assertions tie both name lists to keyof RunStore in each direction and reject a public member the interface does not declare. Each was re-verified by making it fail after the generator was removed. Also drops the knip entry that existed only to treat that script as an entry point.
The sweeper needs to know which run statuses are terminal and cannot import the list, because run-engine depends on run-store rather than the other way round. The copy's comment claimed a parity test kept the two equal. No such test existed, so the claim was false and the copy could drift silently. Drift is not symmetric. A status added to the engine and not the copy makes the sweep treat a finished run as live and never apply its completion expiry. A status removed from the engine and not the copy makes it treat a live run as finished, and that reaps state a run is still using. Verified by removing a status and rebuilding: the test fails and reports seven members against eight.
The forwarders were (...args: any[]): any, so the compiler could not see inside them. A body that called the wrong delegate member, or reordered its arguments, typechecked cleanly. That is not a theoretical gap: it is why a runtime probe existed to catch it, and it is the same shape of hole that let three other defects on this branch pass a green suite. Every member now restates its interface signature and forwards its arguments by name, so both mistakes are compile errors. Verified by making them: a forward to the wrong member produces two type errors, and swapping two arguments produces one. Seven members are overloaded. TypeScript cannot express a single body that satisfies an overload set, so their overloads are declared for callers and their one implementation forwards through a cast. That cast is now the only place the compiler is not checking the forward. The probe shrinks to what is left: those seven casts, a dropped OPTIONAL argument (omitting a trailing tx compiles and silently stops forwarding the transaction), and whether the data property is read live or captured once. Its header states which of those the compiler already covers. Headers on both files now describe what they are rather than that they were once scaffolded.
Typing the forwarders closed the wrong-member and reordered-argument holes but not this one: omitting a trailing OPTIONAL argument still compiles. Two forwarders did exactly that, because the retyping pass read parameter names with a pattern that a preceding inline comment defeated, and both affected parameters happened to be optional and commented. The effects were silent and not small. findLatestExecutionSnapshot stopped applying its tenant scope, so a direct use of the base could read across the environment boundary. upsertWaitpointTag stopped applying its residency hint, so a tag write for a new-database environment would land on legacy. A source-level guard now asserts that every single-signature member forwards exactly the parameters it declares, in order. It reads the interface and the base and compares them, because that property is invisible to the compiler by definition. It carries a vacuity check, so a parse failure fails the suite instead of quietly matching nothing, and that check earned itself immediately by catching a parser that skipped every generic member. Verified: with a parameter dropped again, typecheck reports zero errors and the guard names the member and the missing argument.
A completed waitpoint with no batch index was invisible to every Redis read, so a run resumed from the store lost that wait's result while Postgres still returned it. That is every wait.for, every single triggerAndWait and every token: the engine passes index as batchIndex ?? undefined, so only waits inside a batch carry one. The cause was reading the id set out of the ordered list. That list is the index oracle and its positions ARE the indexes, so it can only ever hold indexed ids, and deduping it yields a set missing exactly the index-less ones. Postgres has no such restriction: its completed-waitpoint join records every id. The cycle key now carries the complete distinct set in its own field, written when the cycle is minted and read back beside the order. The order keeps its meaning and stays index-only. Two tests: one asserting an index-less wait survives a round trip with an empty order, and one asserting the set matches the Postgres join for a mix of indexed and index-less waits. Verified by deriving the set from the order again, which makes the waitpoint vanish. The suites missed this because every earlier case gave each waitpoint an index.
The previous fix stored the complete id set but left three places still deriving it from the ordered list, and the ordered list holds only batch-indexed ids. A carry-forward decided on the order alone. Two DIFFERENT single waits both present an empty order, so they compared equal, the second inherited the first's cycle, and a read returned the wrong waitpoint entirely. The comparison now requires the id set to match as well. The dequeue site built its Redis refs from the ordered list while the delegate connects the complete set in Postgres, so an index-less waitpoint reached Postgres and never reached Redis. Refs are now built from the complete set, with the index taken from the ordered list where the id appears in it. The entry decode derived the set from the order too, which meant getLatest and getById returned an incomplete set. That is the hot read: findLatestExecutionSnapshot hydrates the waitpoint rows from it, so a resume would have fetched no row at all for a single wait. The read scripts now return the stored set alongside the order. Four tests, each verified against its own defect: two consecutive single waits keep separate cycles, a repeated one still carries forward, the dequeue snapshot keeps an index-less id, and the hot read hydrates its row.
A sweep for values derived where they should be read found one more. The hydrated payload left out lastHeartbeatAt entirely, so a Redis-served read returned undefined for it where Postgres returns null. No code writes that column, so null is not a guess: it is the only value Postgres ever holds. The effect was small but constant, on every read served from Redis, and it is the kind of difference a comparator has to either explain or chase. Guarded by comparing the KEY SET of the two payloads rather than their values, so a column omitted by the hydrator fails as a missing key rather than passing as an absent value. Verified by removing the line again: the test names the column. Also covers the timestamp write on both schema variants. updatedAt is declared @updatedat, which Prisma manages, so whether an explicit value survives a create is a property of the client rather than of the schema, and the two variants are separately generated clients. Agreeing declarations were not evidence. Both honour the caller's instant.
An independent pass hunting one shape, a value derived where it should be read, found these. None was reachable from a test that existed. The hot read paid a second Redis call in its most common case. An entry with no wait cycle has no waitpoints by construction, and the hydrator asked the store to confirm that rather than concluding it, on every read of a run that is not resuming from a wait. It now distinguishes the three cases and only asks when it genuinely does not know. decodeWaitpointIds still reconstructed the id set from the ordered list when the stored set was absent. That is the sixth instance of the bug fixed five times, surviving as a fallback. It is unreachable today, because both fields are written by one command, but the reconstruction is lossy by nature and the loss is silent. A missing set beside a non-empty order now reports the entry as not present, which sends the caller to Postgres. The window read checked one liveness anchor where the append script deliberately checks two and explains why. An index lost to eviction while the entry hash survived would have reported an empty hit rather than a miss, so the poll would have returned nothing new for the rest of the run's life while Postgres held the transitions. The wrapped store handle dropped the staging buffer, so a handle taken inside a transaction would have appended before the commit. No caller writes a snapshot through it today. Also restores excess-property checking on the nested snapshot writes. Routing them through a generic helper let a typo'd field name compile and fail at runtime; a concrete parameter type brings the check back at the five sites that pass a fresh literal. Verified: a bogus field is now TS2353.
Two paths reached the same silent hang, and neither had a test. When the store refuses a carried pointer it was still writing the entry, which then became the run's head with no pointer at all. A read of that answers present-with-nothing, and present-with-nothing is precisely the signal that tells the engine's read-repair it does not need to look, so the runner got a waitpoint-less continue and dropped it. Refusing the pointer stays right; the append now mints a fresh cycle from the refs the caller carried, in the same atomic call, so the entry always has a pointer that can be trusted. Refs are optional and only the fallback needs them, so callers that supply none keep the previous behaviour. The second path needs no refusal at all. An entry whose cycle key has gone still carries its pointer, and the read answered empty for it too. That is reachable by eviction and also by the completion expiry, which is applied to every key for a run at one moment but lets them expire independently. Reads now report such an entry as not present, which sends the caller to Postgres, where the join rows still are. The hot read and the window both fall back rather than serve it. Three tests. The refusal is driven at the store, because the decorator cannot reach it on purpose: its probe sees the id set no longer matches and mints a new cycle, so the refusal only happens when the key vanishes between probe and append. Each verified against its own defect.
At that position Postgres holds no snapshot rows, so a run routed away from Redis by the cohort percentage reads nothing at all. The percentage is only meaningful while both stores hold the data. Fixing it in the dial rather than documenting the constraint makes the combination unreachable, instead of leaving three settings that have to agree by convention.
…ecorator-tri-13449 # Conflicts: # internal-packages/run-store/src/index.ts
…edis cluster SCAN carries no key, so a cluster cannot route it: one connection iterates one node's keyspace and then reports a completed cursor. The sweep now fans out over every master, resolved per pass so a failover cannot leave it scanning a stale node list, and reports how many it covered. Rule 2 deletes a whole keyspace when the run lookup returns no row. That lookup partitions ids by residency and reads each store's replica, so an absent row is not proof of absence. Deletion now needs the keyspace to be seen absent in two separate passes, and any run found to exist clears its mark. Both window reads returned the head's waitpoint order without its dangling flag, so a head whose cycle key had expired came back with an empty order rather than falling back to Postgres, losing every position on a batched resume. Also lets both classes take a caller-built client so they can reach a cluster at all, and gives the sweep a deadline and an abort signal so a pass can stop inside its budget instead of being killed mid-cursor.
…ecorator-tri-13449
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
…birth The per-organisation dial is served from a short-lived cache, and on a miss the resolver answered with the deployment-wide position. For a read or a transition that is the right trade. For a birth it is not: residency is decided at birth and is permanent, so a run born during a miss is excluded from the mirror for its entire life and nothing can adopt it later. Observed with an organisation opted in: three runs born back to back were all resident, then after a fourteen minute idle gap the next run was not, because the cache entry had expired. A miss is not a rare event, it is any gap longer than the cache lifetime, so on bursty traffic the first run of every burst was lost and a low-traffic organisation would have lost most of its runs. A birth now waits for the organisation's own value. The wait is bounded: a birth is on the trigger path and a caller may already hold an open transaction, so a slow read gives up and leaves the resolver answering as it did before rather than holding that transaction open. A failed read never fails a trigger. Transitions are unchanged. They stay synchronous and blind to both the dial and the organisation, which is what stops a run changing stores half way through its life.
…t bite Four defects, all in code added earlier in this branch. Dropping a run stopped sweeping wait cycle keys after a run of absent ones. Cycle keys can be sparse, so with the sequence hash already gone and only a high-numbered key alive, the loop stopped before reaching it, and the entry hash is removed in the same call so the sweep could never discover it either. The bounded range is now swept unconditionally. A repair whose append came back as a duplicate returned before marking the keyspace, so it went on serving short windows as whole. A repair runs because an append was lost; whether that one entry had already landed says nothing about the entries either side of it. The read source was recorded before hydration, so one logical read incremented both the Redis and the Postgres series once the hydration fallback landed. It is now recorded once, after hydration decides. A failed primary read of an organisation's dial still cleared the pending flag, reopening the window to a lagging replica that could restore the pre-save value. The load now reports whether it answered. Also: the sweep counter is seeded at zero when the sweep is wired, so the alert asking for no completed passes in 24 hours can fire at all. An absent series matches nothing, which was exactly the case the alert exists for. The seeding is a named function with the seeded outcomes beside it, and a test asserts every outcome the alert rules query is seeded, so neither side can be changed alone. Two tests replaced production methods and no longer do. One observes the store's own read metrics; the other subclasses the store, so every command except the one under test goes through the real path. Each new test was checked by reverting its fix and confirming it fails.
The lowest position stopped new runs joining the mirror but did not take the store off the run path. A transition has to ask whether its own run is resident, the keyspace is the only record of that, so every transition of every run still made one request. Two per cent with a healthy endpoint, four times the run duration with a slow one, for every run, and it did not decay as resident runs drained. Before a deployment has ever enabled the store nothing can be resident: only a birth creates a keyspace and every birth was refused. So the question has one possible answer and asking it is pure cost. A one-way latch records whether the store has ever been enabled, and while it is unset a transition skips the store entirely. Once set it is never cleared automatically, and the lowest position goes back to meaning drain: a run resident from an earlier ramp keeps mirroring, because freezing its head while Postgres moves on makes turning the dial down worse than leaving it alone. The save guard refuses to enable the deployment dial or any per-organisation override until the latch is true. Without that a run born between the two writes would be resident with its transitions skipped, and its head would freeze. Refusing makes the ordering impossible to get wrong rather than merely documented. The latch reads false only when the flag is explicitly false, so a cold or unreadable flag registry reports enabled and can never suppress a transition for a run that is resident. This moves the safe deployment posture from something an operator has to remember to the default. The store can ship with its endpoint configured, off the run path, and every step after that is a flag that converges in seconds rather than a deploy that takes hours.
Comment on lines
+207
to
+208
| if (cache.get(organizationId) !== undefined) { | ||
| return; |
Contributor
There was a problem hiding this comment.
🟡 Remote rollout leaves births behind
After another process changes an organisation dial, warm trusts its cached value until expiry. Births during that window permanently join the previous rollout cohort.
Was this helpful? React with 👍 or 👎 to provide feedback.
Collaborator
Author
There was a problem hiding this comment.
Accurate, and worth stating precisely rather than fixing badly.
The window is the organisation cache lifetime, 30 seconds by default. A process that has a cached value for an organisation trusts it until it expires, so if another process changes that organisation dial, births here follow the old answer for up to that long. Because residency is permanent, those runs stay in the previous cohort for their whole lives.
Not fixed here, because the honest fixes are both worse than the problem at this size. Reading the organisation row on every birth is a query per run on the trigger path. Invalidating across processes needs a broadcast this area does not have, for a value that is already eventually consistent by design.
What bounds it: the window only exists at the moment an organisation is enabled or disabled, not continuously. An operator ramping cohorts can lower the cache lifetime for the duration, which is a flag-free environment value, and the miss path already errs toward the deployment-wide position rather than a stale enable.
Related and now fixed: the same class of bug on a COLD cache was losing runs permanently, because a miss answered with the deployment-wide position and warmed off-path. Births now wait for the organisation's real value, bounded, with the previous behaviour as the fallback.
The latch check took its argument optionally so as not to disturb existing callers. One admin route omitted it and therefore skipped the check entirely, which let a per-organisation override be enabled with the latch unset: those runs would be resident with their transitions skipped, and their heads would freeze. An optional safety argument disables the safety at every caller that forgets it, so it is now required and the compiler enumerates the call sites. There was exactly one. The latch is deployment-wide, like the dial and the hard stop, so an organisation save now strips it rather than reporting success for a setting nothing reads from an organisation row. Bounds on the new numeric configuration. A zero cache lifetime or size is a constructor error, a zero sweep budget truncates every pass so rule 2 can never converge, and a zero command timeout fails every command. Jitter may legitimately be zero. The cluster service pins its image by digest, as the repository requires.
…h preserves the latch
…is fault does not stall sweeps The sweep lock release is a write, so the same Redis brownout that fails a sweep pass also fails the release, and a single best-effort attempt left the lock held for its full TTL (sweep budget plus two hours). One transient blip could pause orphan sweeping across the fleet for hours. The release now retries over a short bounded window, so a Redis that recovers within seconds frees the lock promptly; the TTL remains only as the last-resort backstop.
…s Postgres writes, throw on miss)
…gres is never served
…figured With no snapshot-store Redis host set, the feature must add zero standing cost. Two paths did not honour that: - The org census started its periodic organization.findMany poll at import, gated only on NODE_ENV, so a production process with no host still polled the replica every reload interval. Gate the poll on a host check that is independent of NODE_ENV. - The run store wired the redis-only Postgres-suppression predicate into every PostgresRunStore unconditionally, so every snapshot write ran a mode resolution (and a background per-org replica read on a cold cache). Thread the predicate through buildRunStore and wire it only when configured; unset leaves the store a plain passthrough that writes every snapshot row. Both now route through a single isSnapshotStoreConfigured gate.
Comment on lines
+1122
to
+1125
| // The org-scoped read position, falling back to the global answer when unresolved. | ||
| const effective = this.modeResolver?.readModeFor?.(runId, environmentId) ?? this.mode; | ||
|
|
||
| if (effective !== "redis-read" && effective !== "redis-only") return false; |
Contributor
There was a problem hiding this comment.
🔴 Redis-only rollback loses active state
After redis-only moves to off, readsFromRedis immediately routes resident runs to PostgreSQL. Their snapshots were suppressed there, so engine reads fail.
Prompt for agents
Preserve read residency when snapshotStoreMode or snapshotStoreOrgMode moves down from redis-only. Existing runs have no PostgreSQL snapshot history because PostgresRunStore suppressed those writes, so routing them by the current dial sends them to an empty or stale store. Introduce a per-run residency/cutover rule or require and complete a resynchronization before reads leave Redis. Cover both global and per-organization redis-only rollbacks and the cache-convergence window across pods.
Was this helpful? React with 👍 or 👎 to provide feedback.
| ...rawRequestedFlags | ||
| } = validationResult.data; | ||
|
|
||
| const requestedFlags = withoutOrgForbiddenSnapshotKeys(rawRequestedFlags); |
Contributor
There was a problem hiding this comment.
🔴 Programmatic org rollout freezes snapshots
A v1 organization enable strips the latch but never calls stampSnapshotStoreOrgEverEnabled. The census then suppresses every Redis transition after birth.
Prompt for agents
Make the v1 merge-semantics organization route stamp snapshotStoreOrgEverEnabled inside the existing locked transaction. Apply stampSnapshotStoreOrgEverEnabled to the merged blob before writing it, preserving an existing latch and setting it for any non-off snapshotStoreOrgMode. Refresh the organization census after commit as the v2 route does. Add a route-level test that enables an org through v1 and verifies the latch is stored and transitions are not classified as definitely never enabled.
Was this helpful? React with 👍 or 👎 to provide feedback.
Comment on lines
1006
to
+1018
| if (result.outcome === "forked") { | ||
| this.logger.warn("snapshot append forked", { | ||
| this.logger.error("snapshot append forked", { | ||
| runId: entry.runId, | ||
| snapshotId: entry.id, | ||
| site, | ||
| actualCur: result.actualCur, | ||
| }); | ||
|
|
||
| // A fork is not a race to shrug at: the two compare-and-set sites assert the head, so once it | ||
| // is wrong every later append from them forks too and the mirror is frozen for the rest of the | ||
| // run. The repair re-derives the head from Postgres without asserting `cur`, which is the one | ||
| // operation that clears this, so ask for it rather than only reporting. | ||
| await this.#enqueueRepair(entry); |
Contributor
There was a problem hiding this comment.
🔴 Redis-only forks discard transitions
At redis-only, #recordOutcome accepts forked or missing-keyspace results after Redis writes nothing. PostgreSQL suppression leaves the transition without any snapshot.
Prompt for agents
Treat append outcomes that did not persist the requested transition as fatal when the effective organization mode is redis-only. In particular, forked and skippedNoKeyspace cannot be repaired from PostgreSQL because its snapshot write was suppressed. Propagate a retryable failure instead of returning success or enqueueing the PostgreSQL-based repair. Keep duplicate as successful, and retain the existing repair behavior below redis-only. Add tests for both non-writing outcomes at global and per-org redis-only.
Was this helpful? React with 👍 or 👎 to provide feedback.
Comment on lines
+242
to
+249
| // And only if the primary actually ANSWERED. `load` swallows its own errors, so a rejected | ||
| // primary read used to clear the flag with nothing cached, reopening the window to a lagging | ||
| // replica that would then restore the pre-save value for a full cache lifetime. Staying | ||
| // pending keeps replica refreshes out until a primary read succeeds; the resolver falls back | ||
| // to the deployment-wide position meanwhile, which is the safe answer. | ||
| if (outcome === "failed") { | ||
| return; | ||
| } |
Contributor
There was a problem hiding this comment.
🟡 Failed invalidation wedges org dial
When the post-save primary read fails, primaryPending remains forever and blocks all retries. That process keeps the global fallback for later births.
Prompt for agents
Ensure a failed authoritative reload after invalidate can retry without allowing a lagging replica to overwrite the saved value. Keep replica refreshes blocked, but retain or schedule a bounded primary retry that eventually clears primaryPending only after a successful authoritative answer. Make warm await or trigger that primary retry rather than waiting only for its timeout. Add a test where the first primary read fails and a later warm or refresh successfully loads the saved organization mode.
Was this helpful? React with 👍 or 👎 to provide feedback.
Comment on lines
+1055
to
+1062
| -- So sweep a bounded range unconditionally. A miss-streak early exit was wrong: cycle keys | ||
| -- can be SPARSE, so with seq gone and only wp:10 alive, stopping after a run of absent keys | ||
| -- leaves it behind, and the entry hash is deleted below so the sweep can never find it | ||
| -- either. Every key here shares the {runId} tag, so this stays inside one slot, and the | ||
| -- bound keeps a pathological run from turning a drop into a long script. | ||
| for probe = cycles + 1, cycles + 512 do | ||
| redis.call('DEL', wpKey(probe)) | ||
| end |
Contributor
There was a problem hiding this comment.
🟡 Long-run cycle keys leak permanently
With a missing sequence hash, dropSnapshotRun deletes only cycles 1–512. Longer runs leave undiscoverable cycle keys that no later sweep can remove.
Prompt for agents
Make dropSnapshotRun remove all wait-cycle keys even when the seq hash and its cycle high-water mark are missing. The current fixed probe range cannot guarantee that for runs exceeding 512 cycles, and deleting the entry hash afterward makes leftovers undiscoverable by the orphan sweeper. Use a cleanup strategy that can discover or track the complete cycle-key set while preserving Redis Cluster single-slot behavior and bounded script execution.
Was this helpful? React with 👍 or 👎 to provide feedback.
Comment on lines
+334
to
+339
| singleton("AssertSnapshotStoreBoot", () => { | ||
| assertSnapshotStoreBootFromEnv().catch((error) => { | ||
| logger.error("Snapshot store boot check failed; refusing to start", { error }); | ||
| process.exit(1); | ||
| }); | ||
| return true; |
Contributor
There was a problem hiding this comment.
🟡 Boot checks run after serving begins
assertSnapshotStoreBootFromEnv runs unawaited for up to 15 seconds while HTTP starts. Invalid redis-only configuration can receive traffic before forced exit.
Prompt for agents
Move the snapshot-store boot assertion into an awaited startup phase before the Express listener or worker consumers become ready. Preserve the required ordering after snapshot-store wiring, but do not expose readiness or accept work until the flag wait, connectivity check, and configuration assertions complete. A failing assertion must reject startup directly rather than terminating a process that has already begun serving.
Was this helpful? React with 👍 or 👎 to provide feedback.
Comment on lines
+3
to
+7
| export function cohortMetricLabel( | ||
| organizationId: string | undefined, | ||
| isCohortMember: (organizationId: string) => boolean | ||
| ): string { | ||
| return organizationId !== undefined && isCohortMember(organizationId) ? organizationId : "other"; |
Contributor
There was a problem hiding this comment.
🔍 Organization IDs create metric series
The org attribute contains organization IDs for every active cohort member. Repository telemetry rules prohibit IDs as metric attributes.
Was this helpful? React with 👍 or 👎 to provide feedback.
Comment on lines
+198
to
+209
| export function createOrgModeSource(clients?: { | ||
| primary: OrgModeClient; | ||
| replica: OrgModeClient; | ||
| }): OrgModeSource { | ||
| const primaryClient = (clients?.primary ?? prisma) as OrgModeClient; | ||
| const replicaClient = (clients?.replica ?? $replica) as OrgModeClient; | ||
| // Defaults inline as well as in the schema: this must not throw when a caller supplies a partial | ||
| // env, and an LRU with neither bound set is a constructor error. | ||
| const cache = new LRUCache<string, CachedOrg>({ | ||
| max: env.RUN_ENGINE_SNAPSHOT_STORE_ORG_MODE_CACHE_MAX ?? DEFAULT_CACHE_MAX, | ||
| ttl: env.RUN_ENGINE_SNAPSHOT_STORE_ORG_MODE_CACHE_TTL_MS ?? DEFAULT_CACHE_TTL_MS, | ||
| }); |
Contributor
There was a problem hiding this comment.
🔍 Testable service owns environment configuration
createOrgModeSource reads global environment configuration internally. Webapp rules require testable services to receive configuration through options.
Was this helpful? React with 👍 or 👎 to provide feedback.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Makes the Redis-backed execution-snapshot store reachable from production, off by default.
With no
RUN_ENGINE_SNAPSHOT_STORE_REDIS_HOSTset, nothing is constructed, no connection is opened, no metric series is registered and no job is scheduled, so the store chain behaves exactly as it does today. A fresh self-host is in that state: none of the new variables appear in.env.example, the docker files, or the Helm chart. The dial cannot activate anything on its own either, because construction is gated on the connection rather than on the dial.Once the connection is configured, expect three Redis connections and the orphan sweep running on its cron schedule, at any dial position. The sweep runs at every position on purpose: an operator has to observe a full pass before dual-write starts. Until the dial moves, the keyspace is empty and a pass finds nothing, but it is not nothing.
Stacked on #4765, which builds the store and the decorator. Review that one first.
The dial is a feature flag, not an environment variable
A sustained append failure burns a task attempt on every state transition, so runs can exhaust their retry budget on infrastructure failure rather than task failure. Dialling down is therefore a correctness control, and it cannot wait for a deploy.
Two catalog keys, because they must accept different values:
snapshotStoreMode(global) holds all five positions.snapshotStoreOrgMode(per organisation) holdsoff,dual-writeandcompareonly.Snapshot reads are global, so an organisation at a read position would read state its own writes never created. The narrower enum makes that unrepresentable rather than documented.
ORG_LOCKED_FLAGSturns out to enforce nothing (it is a client-side predicate and no save path consults it), so the line is held the way the mint grace stamps hold it: both organisation routes strip the global key from an incoming payload.The environment keeps what cannot be hot-swapped, plus
RUN_ENGINE_SNAPSHOT_STORE_MODEas the floor used when no flag row exists, so a self-host install still works with no rows at all. No variable falls back to the genericREDIS_*: this is a distinct durable endpoint, and a fallback would silently put execution state on the general-purpose cache.The resolver never queries
The dial is read on every snapshot write, while the per-run lock is held. Seven decorator methods accept a caller-supplied transaction, so this code cannot see a caller's transaction boundary, and an awaited read could land inside someone else's open interactive transaction, on the same connection pool for single-DB and self-host.
So the resolver is synchronous. The global value comes from the existing
globalFlagsRegistry, already an in-memory snapshot read synchronously on the trigger hot path. The per-organisation value comes from a bounded LRU. A miss returns the global answer and warms the cache off-path, so a cold organisation costs no round trip and a control-plane blip cannot fail a state transition.Gating, and the sweep
Construction is gated on the connection, not the dial. The store's options require a constructed Redis store, that store opens its socket in its constructor, and the client factory sets no
lazyConnect, so building it unconditionally would open a doomed localhost connection in every self-host install, developer machine and CI run.The orphan sweep runs as a cron job on the engine's existing worker.
enqueueOncegives no overlap protection (its dedup record is the queue item and the ack deletes it, and nothing extends the visibility timeout), so each pass takes a fenced lock released with a compare-and-delete. A bareDELwould let a pass that overran its own lock delete its successor's.The append-failure hook binds to the existing repair job, sharing the stall watchdog's job id and its
availableAt, so the two compensators can never enqueue two repairs for one run and neither can win a race that changes the delay.Notes for review
Three existing assertions changed, deliberately.
runInTransactionnow always installs the staging facade andforWaitpointCompletionalways wraps its handle: both were conditional on the dial, which cannot work once the dial moves at runtime and a per-organisation value lives in an organisation row. The replacements assert the property that matters, that nothing is appended atoff.@internal/redisgains a cluster-capable client. Cluster mode currently reaches the sweep connection but not the hot path, because the Redis snapshot store still builds its own single-node client; that becomes a one-line change once its options accept a pre-built client.Six guards are verified by reintroducing the defect rather than by observing a pass: the organisation key strip, the module-scope instrument check, the construction gate, the graded boot check, the lock fence and the repair job id each fail when their guard is removed.
Three items are threaded and not yet honoured, because they need #4765's store API: the sweep budget, the abort signal, and
CONFIRM_ORPHAN_AFTER_MS. Cluster mode reaches the sweep connection but not the hot path for the same reason. Each is commented at its site.No changeset and no server-changes entry: the dial defaults off,
@internal/redisis not consumed independently, and nothing user-visible changes yet.