Files
WebHooker/server/lib/queue/consumer.ts
T
翅膀喵andRhenCloud 978c53eeb3 fix(security): harden webhook delivery pipeline and group storage (#11)
* fix(log): preserve structured error details

Error objects lose their useful details when serialized as plain JSON.
Keep the message, stack and cause in structured log output.

* fix(runtime): bound shared in-flight waits

A canceled request can leave its shared promise pending forever.
Bound joins and coordinate a single replacement so later requests
make progress without multiplying upstream work.

* fix(bot): reject oversized callback bodies

Reject declared oversized bot callbacks before reading their bodies.
This keeps unauthenticated requests from wasting worker resources.

* test(d1): exercise real sqlite transactions

Exercise real foreign keys, transactions and conditional statements
instead of mocks that cannot reproduce lost updates or cascades.

* fix(storage): preserve claims across kv fallback

Only missing tables justify KV fallback; other failures must surface.
Remember fallback writes so migration to D1 cannot forget claims
and replay completed work. Share marker checks without extra writes.

* fix(filter): use isolated linear-time patterns

Backtracking patterns can exhaust worker CPU on untrusted payloads.
Use linear-time matching and separate bounded cache namespaces so
glob and regex entries cannot corrupt each other.

* fix(filter): read each event branch correctly

Deployment and check branches live in event-specific nested fields.
Read those fields and remove only the actual leading ref prefix.

* fix(filter): read only own payload properties

Inherited object properties are not webhook payload data.
Read own properties only so JSONPath cannot traverse prototypes.

* fix(filter): validate executable condition trees

Malformed trees and unsupported patterns must not silently turn an
exclude rule into a pass. Validate the shape and only the patterns
the evaluator actually executes.

* fix(i18n): bound and cache translation loads

Repeated misses and simultaneous override loads spend KV reads for
the same language. Cache confirmed misses and share bounded loads
without caching read failures as missing translations.

* i18n(bot): explain command and account failures

Explain account linking, command failures and skipped delivery
outcomes consistently in both supported languages.

* fix(github): bound and reuse app token fetches

Repeated token exchanges waste requests and JWT work for every event.
Reuse valid installation tokens, share refreshes and back off briefly
after failures while keeping every shared wait bounded.

* fix(github): restrict enrichment request urls

Webhook-provided URLs must never receive an installation token on
another host. Restrict enrichment to GitHub repository API paths,
reject redirects and bound optional requests.

* fix(formatter): preserve invalid code points

Out-of-range numeric entities can throw while formatting a delivery.
Keep invalid code points verbatim instead of aborting the message.

* fix(formatter): preserve bounded markdown text

Code spans, fences, indentation and autolinks are Markdown content,
not HTML tags to strip. Preserve them while bounding cleanup input
and reusing results across routes for the same event.

* fix(formatter): strip only leading ref prefixes

A short branch or tag may itself contain refs/heads or refs/tags.
Strip one leading full-ref prefix at the caller and keep short names
unchanged when generating labels and links.

* fix(gitea): link push authors to their forge

A Gitea author is not the same-named GitHub account.
Build author profile links from the originating forge.

* fix(formatter): render deployment state values

The deployment status object stringifies as [object Object].
Render its actual state in the status field.

* fix(formatter): read security severity levels

The scanner severity is not the security severity of an alert.
Prefer the security level so critical and high alerts get the right
label and color.

* fix(formatter): hide actions on other forges

Action buttons call the GitHub API as the linked user.
Hide them for other forges to avoid operating on a same-named
GitHub repository.

* fix(check): isolate updates by check suite

Same-named checks from different suites must not edit one message.
Include the suite or app identity and retain the former key for
controlled adoption of existing tracked messages.

* fix(logs): scope queries before limiting results

Filter authorized groups in SQL before LIMIT and aggregate only
visible rows. Preserve current group attribution for late writes,
surface read failures and combine related totals in one scan.

* perf(d1): index pruning and failed send queries

Pruning and failure views should not scan all historical rows.
Index the high-value access paths without adding a write-heavy
index for infrequent delivery-detail lookups.

* fix(discord): bound individual api requests

An unbounded API request can outlive an update-chain lock.
Bound each request so another holder cannot race a late edit.

* fix(discord): defer long rate-limit waits

Long rate-limit sleeps block delivery progress inside a worker.
Return a retryable 429 for the queue when the wait exceeds the bound.

* fix(discord): encode channel path segments

Stored channel and message ids are path segments, not URL paths.
Encode them so aliases cannot address another Discord resource.

* fix(discord): reuse nonces after lost responses

A successful create may lose its response. Reuse a stable nonce
with enforce_nonce so retries recover that message instead of
posting duplicates or accepting an unreadable success.

* fix(discord): constrain embeds and role mentions

Respect combined embed and component limits as well as field caps.
Restrict allowed mentions to validated route roles so payload text
cannot notify everyone.

* fix(discord): resume bounded command syncs

Application changes, command rollbacks and large guild lists can
leave commands stale indefinitely. Track the current hash per app
and resume bounded paginated work without needless KV writes.

* fix(github): verify installation management

An installation id is user-controlled, and repository access is not
installation management. Verify the owning account or an active
organization admin using the user token and read:org scope.

* fix(github): resolve immutable member ids

A login can be renamed and reclaimed before its first sign-in.
Resolve the current immutable user id when membership is granted
so a future holder of the login cannot inherit access.

* fix(auth): revoke stale sessions and tokens

Revoked or rotated credentials must stop working despite stale KV
replicas. Track the live token and revocations in D1 and preserve
immutable identity bindings rather than trusting a recycled login.

* fix(auth): bind sessions and tokens to identity

Automatic retries can repeat non-idempotent user actions.
Disable those retries and expose a write hook so bot handlers can
record completion immediately before the GitHub write.

* fix(discord): defer and deduplicate interactions

Acknowledge slow actions before Discord times out and deduplicate
interaction writes. Preserve comment text and reject edits whose
source changed after the modal opened.

* fix(telegram): always verify webhook secrets

Webhook verification must not silently disable itself when the
optional secret is absent. Derive a stable secret from the bot
token and reject callbacks when verification is unavailable.

* fix(telegram): bound individual api requests

Bound individual platform calls so they cannot outlive the lock
protecting a message update or stall a whole delivery invocation.

* fix(telegram): honor nested rate-limit delays

Telegram places retry_after inside parameters, not at the top level.
Read that value and defer long waits to the delivery queue.

* fix(telegram): report unreadable send results

An unreadable successful response may already have sent the message.
Report an explicit uncertain result instead of retrying a create
without an idempotency key or pretending it succeeded.

* fix(telegram): keep editable messages as text

Guessing photo versus text from message content makes later edits
use the wrong API and a smaller caption budget. Always send text
and use the avatar as a link preview.

* fix(telegram): render hard breaks as newlines

Telegram HTML does not support br tags. Render Markdown hard breaks
as plain newlines so otherwise valid messages are accepted.

* fix(telegram): clamp visible html characters

Telegram counts visible characters, not HTML source bytes.
Truncate without splitting tags, entities or surrogate pairs.

* fix(telegram): honor the author emoji toggle

The author marker must obey the group emoji preference just like
formatter-added decoration.

* fix(telegram): keep completed commands claimed

Redelivery after a completed write must never post another comment.
Keep completion claims despite reply failures, scope updates per bot
and only issue account-login links in private chats.

* fix(feishu): bound individual api requests

Bound individual platform calls so they cannot outlive the lock
protecting a message update or stall a whole delivery invocation.

* fix(feishu): classify retryable api failures

Business error codes can be transient even in HTTP 200 responses.
Classify token and rate-limit errors for retry and defer long waits.

* fix(feishu): retry sends with stable uuids

An unreadable response does not prove a send failed.
Reuse the same uuid for retries so recovering the result cannot
create a second Feishu message.

* fix(feishu): refresh and evict tokens safely

Separate app token caches, stop reusing refused tokens and share
refreshes. Cache failures briefly and keep valid tokens in memory
when a best-effort KV cache write fails.

* fix(feishu): use the card update endpoint

Delayed card updates use the card update endpoint; the config
endpoint cannot update an interaction response.

* fix(feishu): neutralize untrusted mentions

Forge text is untrusted and can contain at-all tags.
Neutralize mentions in card bodies and footers before rendering.

* fix(feishu): authenticate and dedupe callbacks

Authenticate signed or token-verified callbacks before dispatching
commands. Decode encrypted envelopes and handle card actions using
the correct fields without claiming irrelevant chat events.

* fix(target): normalize actual destinations

Different target spellings can name the same actual destination.
Normalize channel, thread and topic identities for ownership checks
and per-route duplicate suppression.

* fix(config): transact per-group configuration

Whole-table writes and cached snapshots can overwrite other groups
or resurrect removed configuration. Use per-group guarded writes,
one-time identity migration and explicit unreadable-row failures.

* fix(group): preserve identity through lifecycle

Group names are mutable and reusable, so they cannot own durable
state. Preserve immutable identities across lifecycle operations and
commit invite use with membership under transaction-time guards.

* fix(oauth): confirm account links and installs

A login link or installation id is not proof of account ownership.
Verify installation management, bind browser state and require an
explicit account-link confirmation before granting chat access.

* fix(queue): fence leases and record target outcomes

Retries must remember completed targets and distinguish active
lease holders. Fence state updates, preserve permanent outcomes and
reject missing parked payloads instead of dispatching empty data.

* fix(dispatch): isolate and resume tenant sends

Check tenant sources before filters and never dispatch orphan routes.
Record each target outcome, isolate update chains by forge and group
identity, and keep stale events from overwriting newer messages.

* fix(queue): bound batches and resume deliveries

A queue invocation has shared query and subrequest limits.
Use a small batch and shared send budget, renew fenced leases and
resume only unfinished work without exhausting unstarted messages.

* fix(admin): guard revisions and tenant targets

Read-time authorization alone cannot protect concurrent edits.
Guard revisions, current roles and destination ownership in the
write transaction, and use per-group endpoints for isolated changes.

* fix(webhook): confirm only accepted deliveries

Acknowledging before queue acceptance permanently loses deliveries.
Confirm only accepted work, release failed claims, scope replay
identities and strip tenant-supplied installation authority.

* fix(storage): prune abandoned parked payloads

Oversized payloads can remain after interrupted enqueue attempts.
Prune objects beyond their retry lifetime without deleting payloads
that a current delivery may still need.

* fix(console): expose api conflict statuses

Callers need the HTTP status to distinguish stale edits from other
failures and reload current data after a conflict.

* refactor(console): model group revisions

Carry server revisions explicitly so console writes can reject
changes based on an outdated group snapshot.

* fix(console): send loaded route revisions

Send the revision actually loaded with the route list and adopt the
server response after a save. Reload on conflict instead of silently
overwriting another editor.

* fix(console): guard fragment saves by revision

Fragments share a group version with other configuration writes.
Submit the loaded revision and reload on conflict to preserve
concurrent edits.

* fix(console): save groups with guarded writes

Whole-list saves can overwrite unrelated groups. Use per-group
revision-aware writes, retain rename results and confirm deletion
on the exact group, routes and version shown to the user.

* fix(console): reject edits of stale routes

A conflict reload can leave an older route editor open.
Compare its original value before saving so it cannot overwrite
the newly loaded route.

* fix(console): show overview loading failures

Failed overview requests must not look like an empty dashboard or
a successful refresh. Show the reason while retaining prior data.

* fix(console): explain rejected install bindings

Explain why an installation binding was refused instead of silently
returning to the console with no visible outcome.

* fix(console): allow selecting a group named all

The all sentinel collides with a valid group id.
Use a value that cannot be a group id so that group stays selectable.

* fix(console): discard stale log detail responses

An older detail request can finish after a newer selection or close.
Discard it so the dialog never shows the previously clicked record.

* fix(console): retain legacy group owners

Groups predating member roles store owners in adminIds.
Load those as owners so adding a member never drops all old owners.

* fix(console): lock member edits during saves

Edits made during a save are overwritten by its result.
Disable member controls and guard handlers until the write settles.

* fix(console): show invite loading failures

Failed invite loads must not look like no pending invites.
Show the error and discard responses from a previously shown group.

* fix(console): retain legacy discord log targets

A stored log target without platform means Discord.
Keep that meaning in the editor instead of clearing the target.

* fix(console): retain undisplayed forge sources

An editor that only displays some providers must preserve the others.
Keep stored custom sources when saving unrelated group settings.

* fix(console): lock the group editor while saving

Prevent edits while a group save is in flight and make super-only
installation settings readonly instead of silently ignoring edits.

* fix(console): lock route edits and show failures

Lock the route editor while saving and show fragment load failures
with a retry action rather than presenting a false empty list.

* fix(config): expose the legal contact override

Declare the public runtime key so the documented environment
override reaches legal pages in deployed Nuxt builds.

* build(deploy): migrate d1 before deployment

Apply pending D1 migrations before deploying code that requires
them. Stop deployment when migration fails instead of running
against an incomplete schema.

* docs(worker): align guides with runtime behavior

Synchronize both language guides, API references and examples with
the implemented authorization, storage and delivery behavior.
Document migration requirements and accepted operational limits.

* fix(storage): repair orphaned KV routes

Repair orphaned KV routes before serving KV-only configuration.

* fix(security): fully escape forge HTML

---------

Co-authored-by: RhenCloud <i@rhen.cloud>
2026-10-24 16:19:06 +08:00

413 lines
17 KiB
TypeScript

import type { Config, Env, Group, WebhookEvent, WebhookProvider } from "../types";
import { dispatchEvent, type DispatchOptions } from "../core/dispatch";
import { loadConfigSnapshot, loadGroupFresh } from "../config";
import { groupAliasChain } from "../web/group-lifecycle";
import { log } from "../lib/log";
import { recordSendBatch } from "../lib/send-log-batch";
import {
DELIVERY_DLQ,
PayloadUnavailableError,
type DeliveryMessage,
type DispatchSummary,
acquireDelivery,
classifyDelivery,
finishLease,
getDeliveryState,
renewLease,
type DeliveryStatus,
deliveryStateKey,
discardPayload,
hadPermanentFailure,
isRetryableFailure,
markDeliveryDead,
markTargetsDone,
resolvePayload,
retryDelay,
targetDoneChecker,
} from "./delivery";
type Dispatch = (
config: Config,
event: WebhookEvent,
env: Env,
groups?: Group[],
options?: DispatchOptions,
) => Promise<DispatchSummary>;
export async function handleQueueBatch(
batch: MessageBatch<DeliveryMessage>,
env: Env,
dispatch: Dispatch = dispatchEvent,
budgetMs = BATCH_BUDGET_MS,
): Promise<void> {
// Shared by the whole batch: the limits it guards (D1 queries and
// subrequests) are per invocation.
const budget: BatchBudget = { deadline: Date.now() + budgetMs, sendsLeft: BATCH_SEND_LIMIT };
for (const message of batch.messages) {
const body = message.body;
try {
if (batch.queue === DELIVERY_DLQ) {
if (await markDead(env, body)) message.ack();
// Another copy holds a live lease and may still need the payload.
else message.retry({ delaySeconds: BUSY_RETRY_SECONDS });
continue;
}
// Out of time or sends before it started (earlier messages took them):
// back to the queue for a later invocation — a retry (one more queue
// read), not a copy (a write and a delete more). Waiting is not a
// failure, though: a delivery never started whose retry would use up
// its last one (sending it to the DLQ untried) is copied instead. One
// that was started has real failures among its attempts; a copy would
// start their count over, so it keeps the retry.
if (spent(budget)) {
if (message.attempts > QUEUE_MAX_RETRIES && env.QUEUE && (await neverStarted(env, body))) {
try {
await env.QUEUE.send(body, { delaySeconds: CONTINUE_DELAY_SECONDS });
message.ack();
continue;
} catch (err) {
log.warn({ deliveryId: body.deliveryId, err }, "Failed to queue a delivery for later");
}
}
message.retry({ delaySeconds: CONTINUE_DELAY_SECONDS });
continue;
}
await processMessage(env, body, message, dispatch, budget);
} catch (err) {
// One message's storage failure must not fail (and replay) the batch.
log.error({ deliveryId: body.deliveryId, err }, "Queue message handling failed");
message.retry({ delaySeconds: retryDelay(message.attempts) });
}
}
}
/** Renew a held lease when it is older than this (the lease lasts five minutes). */
const LEASE_RENEW_MS = 60_000;
const FINAL: ReadonlySet<DeliveryStatus> = new Set(["delivered", "failed", "dead"]);
/** Wait for an unexpired lease of another consumer: its full length plus a margin. */
const BUSY_RETRY_SECONDS = 310;
/**
* Sending time of one batch: a consumer invocation may run 15 minutes, and a
* target takes under a minute at worst (bounded platform calls), so sending
* stops in time to record what was done.
*/
const BATCH_BUDGET_MS = 12 * 60_000;
/**
* Targets one batch sends at most: each costs D1 queries (its ledger record,
* an update chain's lock and tracking) and subrequests, and one invocation
* may make only 50 of each on Workers Free (with ~7 D1 queries per delivery
* besides; batches are small, see wrangler.jsonc). A message left unstarted
* goes back to the queue (a retry: one queue read); a message cut short is
* continued by a new one only when it made progress.
*/
const BATCH_SEND_LIMIT = 20;
/** What a batch may still spend: its time and its sends. */
interface BatchBudget {
deadline: number;
sendsLeft: number;
}
function spent(budget: BatchBudget): boolean {
return Date.now() >= budget.deadline || budget.sendsLeft <= 0;
}
const CONTINUE_DELAY_SECONDS = 5;
/** The delivery queue's `max_retries` (`wrangler.jsonc`). */
const QUEUE_MAX_RETRIES = 5;
/** No attempt has taken the delivery's lease yet (its attempts were all waits). */
async function neverStarted(env: Env, body: DeliveryMessage): Promise<boolean> {
const key = deliveryStateKey(body.provider, body.groupId, body.deliveryId, body.groupInstance);
const state = await getDeliveryState(env, key);
return state == null || state === "pending";
}
/** Clean up after a terminal state; a failure must not reopen the delivery. */
async function discardQuietly(env: Env, body: DeliveryMessage): Promise<void> {
await discardPayload(env, body).catch((err) =>
log.warn({ deliveryId: body.deliveryId, err }, "Failed to discard a parked payload"),
);
}
async function processMessage(
env: Env,
body: DeliveryMessage,
message: Message<DeliveryMessage>,
dispatch: Dispatch,
budget: BatchBudget,
): Promise<void> {
const key = deliveryStateKey(body.provider, body.groupId, body.deliveryId, body.groupInstance);
if (await finishedUnderLegacyKey(env, body)) {
message.ack();
return;
}
const lease = await acquireDelivery(env, key);
if (lease.status === "finished") {
message.ack();
return;
}
if (lease.status === "busy") {
// Another consumer holds this delivery (or a holder vanished): one retry
// after the lease surely expired, instead of spending the retry budget
// polling a lease that is still valid.
message.retry({ delaySeconds: BUSY_RETRY_SECONDS });
return;
}
let token = lease.token;
let renewedAt = Date.now();
let leaseLost = false;
/** A renewal failed with an unknown outcome: stop sending, but never assume a new holder. */
let leaseUncertain = false;
let renewing: Promise<boolean> | null = null;
/** The batch's budget ran out: the rest is left to a continuation message. */
let outOfTime = false;
/** End this attempt with `status`; acks instead when the lease was lost. */
const finish = async (status: DeliveryStatus, then: "ack" | "retry"): Promise<void> => {
const owned = !leaseLost && (await finishLease(env, key, token, status));
if (owned && then === "retry") {
message.retry({ delaySeconds: retryDelay(message.attempts) });
return;
}
if (!owned && !leaseLost) {
// The outcome could not be written, yet no new holder is known (our own
// renewal may have gone through unseen, or the lease lapsed unnoticed):
// come back once any lease surely expired; the state then decides.
message.retry({ delaySeconds: BUSY_RETRY_SECONDS });
return;
}
// Lost lease: the new holder owns the delivery (and its outcome).
message.ack();
if (owned && FINAL.has(status)) await discardQuietly(env, body);
};
let summary: DispatchSummary;
// Targets recorded as sent the moment they were sent (see `afterTarget`).
const recorded = new Set<string>();
try {
const payload = await resolvePayload(env, body);
const event: WebhookEvent = {
event: body.event,
payload,
deliveryId: body.deliveryId,
provider: body.provider as WebhookProvider,
installationId: body.installationId,
};
// Read fresh, not from this isolate's cache: a group (or route) created
// moments ago elsewhere must not be taken for missing — for a tenant
// delivery that verdict is final.
const { config, groups } = await loadConfigSnapshot(env, { fresh: true });
// A tenant delivery belongs to the group instance that received it — under
// whatever id it has now — and never to another group that took over the
// id since. A message from before instances belongs to the group that had
// its id back then (`legacyId`), also when that group was renamed before
// instances existed (followed through the rename aliases) — never to a
// group created since or renamed to the id.
let tenantId = body.groupId;
if (tenantId && body.groupInstance) {
tenantId = groups.find((g) => g.instance === body.groupInstance)?.id;
} else if (tenantId) {
const owner = (id: string): string | undefined => groups.find((g) => g.legacyId === id)?.id;
let found = owner(tenantId);
for (const id of found ? [] : (await groupAliasChain(env, tenantId)).slice(1)) {
found ??= owner(id);
}
tenantId = found;
}
if (body.groupId && (!tenantId || !groups.some((g) => g.id === tenantId))) {
// A group row that cannot be read may be the tenant: retry until it is
// repaired rather than give the delivery up as one of a deleted group.
if (config.unreadableGroups?.length) {
log.error(
{ deliveryId: body.deliveryId, unreadable: config.unreadableGroups },
"Tenant unresolved while some group configuration is unreadable",
);
await finish("retrying", "retry");
return;
}
// The group was deleted: nothing can receive the event any more.
log.warn({ deliveryId: body.deliveryId, groupId: body.groupId }, "Tenant group is gone");
await finish("failed", "ack");
return;
}
const options: DispatchOptions = {
tenantId,
// Always consult the ledger: another queue message of the same delivery
// (a sender redelivery) may already have sent some targets.
skipTarget: targetDoneChecker(env, key),
// Keep the lease alive during a long dispatch; once it is lost, stop
// sending (the new holder takes over the remaining targets).
beforeTarget: async () => {
if (leaseLost || leaseUncertain || outOfTime) return false;
if (spent(budget)) {
outOfTime = true;
return false;
}
budget.sendsLeft--;
if (Date.now() - renewedAt < LEASE_RENEW_MS) return true;
// Routes run in parallel: they share one renewal. Two renewals with
// the same token would race, and the loser would take its own
// successful renewal for a lost lease.
renewing ??= (async (): Promise<boolean> => {
try {
token = await renewLease(env, key, token);
renewedAt = Date.now();
leaseLost = token == null && lease.token != null;
return !leaseLost;
} catch (err) {
log.warn({ deliveryId: body.deliveryId, err }, "Lease renewal failed");
leaseUncertain = true;
return false;
} finally {
renewing = null;
}
})();
return renewing;
},
afterTarget: async (targetKey) => {
await markTargetsDone(env, key, [targetKey]);
recorded.add(targetKey);
},
};
const scoped = tenantId
? { ...config, routes: config.routes.filter((r) => r.groupId === tenantId) }
: config;
summary = await dispatch(scoped, event, env, groups, options);
} catch (err) {
if (err instanceof PayloadUnavailableError) {
// Nothing left to dispatch in this message: drop it, loudly, without
// retrying forever or pretending the delivery succeeded. The delivery
// stays open (not final): the sender's redelivery, with its own payload,
// can still deliver it.
log.error({ deliveryId: body.deliveryId, err: err.message }, "Queued payload unavailable");
// On record where admins look (the group's send log and failure
// metrics): the event needs a redelivery from its sender.
await recordSendBatch(env.DB, [
{
ts: Date.now(),
routeId: "",
groupId: body.groupId,
groupInstance: body.groupInstance,
event: body.event,
target: "",
ok: false,
error: `${err.message}; redeliver the webhook to send it`,
errorCode: "PAYLOAD_MISSING",
deliveryId: body.deliveryId,
},
]);
await finish("retrying", "ack");
return;
}
log.error({ deliveryId: body.deliveryId, err }, "Queue delivery failed");
await finish("retrying", "retry");
return;
}
const { failed, retryable } = classifyDelivery(summary);
// Record finished targets first, whatever comes next: if a later write
// fails and the message is retried, these targets are never sent again.
const permanent = summary.failures.filter((f) => !isRetryableFailure(f));
try {
await markTargetsDone(
env,
key,
[...(summary.succeeded ?? []), ...permanent.flatMap((f) => f.key ?? [])].filter(
(k) => !recorded.has(k),
),
{ permanentFailure: permanent.length > 0 },
);
} catch (err) {
// The outcome is not on record: retry (giving the lease back now) rather
// than finish the delivery with a forgotten permanent failure.
log.error({ deliveryId: body.deliveryId, err }, "Failed to record delivery targets");
await finish("retrying", "retry");
return;
}
if (failed && retryable) {
// Stopped for time, not for a failure: continued by a new message (sent
// before the lease is given back; a lease still held only delays it) —
// only when this attempt finished targets. One that made no progress
// (its targets kept failing until time ran out) is an ordinary retry and
// counts against the retry limit, so it cannot loop forever.
const progressed = (summary.succeeded?.length ?? 0) > 0 || permanent.length > 0;
if (outOfTime && progressed && !leaseLost && env.QUEUE) {
try {
await env.QUEUE.send(body, { delaySeconds: CONTINUE_DELAY_SECONDS });
await finish("retrying", "ack");
return;
} catch (err) {
log.warn({ deliveryId: body.deliveryId, err }, "Failed to queue the rest of a delivery");
}
}
await finish("retrying", "retry");
return;
}
// Targets that failed for good in an earlier attempt still count.
const anyPermanent = failed || (await hadPermanentFailure(env, key));
await finish(anyPermanent ? "failed" : "delivered", "ack");
}
/**
* The delivery already finished under the state key used before deliveries
* were scoped by `@global` / group instance (`delivery-state:{provider}:
* {groupId | "global"}:{id}`, kept 7 days): a sender's redelivery of an event
* handled before the upgrade is then not sent again. The old key is only
* trusted where its owner is certain — for a tenant, the group that had the
* id back then (its `legacyId`); for the global endpoint, while no group named
* "global" (which used the same key) exists or was deleted or renamed away
* recently.
*/
async function finishedUnderLegacyKey(env: Env, body: DeliveryMessage): Promise<boolean> {
// The global endpoint's old key was shared with any tenant once named
// "global" — even one deleted before the upgrade, which nothing records:
// its owner is never certain, so it is never trusted (a redelivery is sent
// again rather than an event dropped). Nothing to read for it, then.
if (!body.groupId || body.groupId === "global") return false;
// The store's one-time setup (seeding from KV, which also turns on reading
// older KV state) must have run before state is read: one group row reads
// it in, not the whole configuration (read below only when it matters).
const { group } = await loadGroupFresh(env, body.groupId);
// The message's own group, still under this id, with no older id of the
// same name (a group created since): nothing it could have finished under
// the old key, so that key is not read.
if (
body.groupInstance &&
group?.instance === body.groupInstance &&
group.legacyId !== body.groupId
) {
return false;
}
// Custom delivery ids got a namespace since (`d:`; `n:` for nonce
// identities, which did not exist before) — in messages queued since. A
// message queued before carries the id as the sender sent it, prefix-like
// text included.
const namespaced = body.provider === "custom" && (body.stateVersion ?? 0) >= 2;
if (namespaced && body.deliveryId.startsWith("n:")) return false;
const id =
namespaced && body.deliveryId.startsWith("d:") ? body.deliveryId.slice(2) : body.deliveryId;
const legacy = `delivery-state:${body.provider}:${body.groupId ?? "global"}:${id}`;
const state = await getDeliveryState(env, legacy);
if (!state || !FINAL.has(state)) return false;
// (A tenant named "global" shared this very key with the global endpoint:
// refused above.)
const { groups } = await loadConfigSnapshot(env, { fresh: true });
const owner = body.groupInstance
? groups.find((g) => g.instance === body.groupInstance)
: groups.find((g) => g.legacyId === body.groupId);
return owner?.legacyId === body.groupId;
}
/**
* Settle a message that left the DLQ; false while another copy of the
* delivery holds a live lease (its state and payload are left alone then).
*/
async function markDead(env: Env, body: DeliveryMessage): Promise<boolean> {
const key = deliveryStateKey(body.provider, body.groupId, body.deliveryId, body.groupInstance);
// A delivery another message of it finished (or is sending) keeps its state.
if ((await markDeliveryDead(env, key)) === "busy") return false;
await discardPayload(env, body);
return true;
}