Aller au contenu
Français

BoundedMailbox

Ce contenu n’est pas encore disponible dans votre langue.

Defined in: src/mailbox/BoundedMailbox.ts:15

Mailbox with a fixed upper bound on queued user messages. Policy for what happens when a message arrives on a full mailbox is configurable.

T = unknown

new BoundedMailbox<T>(options): BoundedMailbox<T>

Defined in: src/mailbox/BoundedMailbox.ts:19

BoundedMailboxOptions

BoundedMailbox<T>

DroppingMailbox<T>.constructor

readonly deadLetterDrops: boolean

Defined in: src/mailbox/DroppingMailbox.ts:71

See DropReportingMailbox.deadLetterDrops. Fixed at construction because the cell reads it once, when it registers its observer: a switch a running mailbox could flip would mean re-reading it per dropped message, on the path this class exists to keep cheap.

DroppingMailbox.deadLetterDrops


droppedCount: number = 0

Defined in: src/mailbox/DroppingMailbox.ts:63

Number of messages dropped by the overflow policy — useful for metrics.

DroppingMailbox.droppedCount

get size(): number

Defined in: src/internal/Mailbox.ts:573

Number of pending user messages.

number

DroppingMailbox.size


get suspended(): boolean

Defined in: src/internal/Mailbox.ts:292

boolean

DroppingMailbox.suspended

dequeueSystem(): Envelope<unknown> | undefined

Defined in: src/internal/Mailbox.ts:562

Envelope<unknown> | undefined

DroppingMailbox.dequeueSystem


dequeueUser(): Envelope<T> | undefined

Defined in: src/internal/Mailbox.ts:467

Envelope<T> | undefined

DroppingMailbox.dequeueUser


drainSystem(): Envelope<unknown>[]

Defined in: src/internal/Mailbox.ts:591

Envelope<unknown>[]

DroppingMailbox.drainSystem


drainUser(): Envelope<T>[]

Defined in: src/internal/Mailbox.ts:587

Drain all user messages; returns them so the caller can forward to dead letters.

Materialises a fresh array rather than handing out the backing store — a ring is not a dense array, so there is nothing to hand out. The allocation is real but it is on the termination path, where the caller (ActorCell) only iterates the result once.

Envelope<T>[]

DroppingMailbox.drainUser


enqueue(env): void

Defined in: src/mailbox/BoundedMailbox.ts:48

A switch rather than a match, deliberately.

The pattern-matching convention governs dispatch of an incoming message, event or command; overflow is none of those — it is a configuration value fixed at construction, and this is the shape the codebase already uses for a closed string-literal union (see decodeCrdt in crdt/DistributedData.ts, which documents itself as the reference).

It matters here because of when this branch runs. A matcher plus one closure per arm was being built for every message that arrived at a full mailbox — that is, once per shed message, at the exact moment the system is already past its capacity and least able to afford it (#974). Exhaustiveness is not lost, only moved: the never assignment below fails to compile if a fourth policy is added without an arm here.

Envelope<T>

void

DroppingMailbox.enqueue


enqueueSignal(env): void

Defined in: src/mailbox/BoundedMailbox.ts:101

A death notification is queued whatever the bound says — see Mailbox.enqueueSignal.

Straight to the base queue, past the capacity check, and that is the whole override: every one of the three policies destroyed the notification otherwise, each in its own way (#729). drop-head evicted whatever sat at the front, drop-new discarded the notification on arrival — the more likely of the two, since a Terminated arrives late relative to the flood that filled the queue — and reject was worse than either: it threw MailboxFullError synchronously on the sender’s stack, and the sender is the dying cell’s own watcher-notify loop. That throw escaped finalizeTermination mid-loop, so the remaining watchers went unnotified, the parent was never told the child had stopped, and terminate() never settled. reject is also this class’s constructor default, so the documented withMailbox(() => new BoundedMailbox({ capacity: n })) shape reached it without naming it.

Envelope<T>

void

DroppingMailbox.enqueueSignal


enqueueSystem(env): void

Defined in: src/internal/Mailbox.ts:463

Queue a framework SystemCommand on the lane that overtakes all user traffic.

This lane is deliberately uncapped, and no bound reaches it — including a subclass’s own. It writes straight to the private system ring rather than delegating, which is the deliberate opposite of enqueueSignal two methods up: a signal rides the user lane and therefore has to go through whatever store a subclass keeps, while a system command has no business in that store at all. So overriding enqueue to shed load bounds the user queue and leaves this one alone, and neither BoundedMailbox nor PriorityMailbox overrides this method. A capacity an operator tunes against measured heap is a user-queue capacity; say so when you document one.

The reason is the one Envelope.undroppable gives one layer up, at full strength: supervision and lifecycle must never be dropped. A discarded failure leaves an actor suspended with nobody coming to decide about it, a discarded childTerminated leaks a child registration forever, and a discarded terminate strands the cell — none of which the framework has a second copy to re-send, and all of which fail silently in the direction of a wedged actor rather than a lost message. Nor does the drain loop in ActorCell.run carry a per-turn budget, for the matching reason: a budget would interleave a supervision decision with user traffic that decision exists to gate, and reordering supervision is a worse trade than the user-message starvation an unbounded drain admits.

What bounds it instead is the producer, and that is the property to re-check before adding a system-message source. Every caller today is one envelope per event that already costs the node an actor: one create per cell and one terminate per stop, one childTerminated per child that finishes stopping, and per supervision decision one failure to the parent plus whatever it fans out over the failing subtree — suspend, then resume, recreate or terminate once the directive is known. One receiveTimeout per armed timer that expires, and one watchNotify per Terminated that a remote watch brings home (#918) — a watch this node’s own actor registered, answered once — complete the list.

watchNotify was for a long time the declared variant deliberately not on it, because nothing emitted one; putting it there as “one per watched death” was this JSDoc’s own first mistake. It still is not how a local death travels: ActorCell._notifyWatcher hands a local watcher its Terminated through postSignalEnvelope, which is the user lane — exempt from the bound (see Envelope.undroppable) and still behind everything already told to that watcher, because that ordering is what death watch documents. Watching a thousand local actors adds nothing to the depth here. What RemoteWatcher enqueues is the one case the arm was kept wired and exempt for (#729): a death this node did not witness, which it turns into the branded Terminated and re-posts on the user lane. The ratio holds for it too — a peer can produce at most one per watch this node opened, an unsolicited watch-terminated matches no entry and enqueues nothing, and the entry itself was paid for by the watcher’s registration.

That test is about volume, not about who can reach the method, which is why the sharding path does not change the answer even though it looks like it should. A remote peer addressing entities in a shard really does drive this queue: dispatchEnvelope only ever reaches a tell on the user lane, but the Shard actor handling that user message spawns and watches an entity, and every entity it spawns eventually lands a childTerminated here at passivation or handoff. “A cluster peer cannot invoke it” is therefore true of the call site and beside the point for a queue depth.

It needs no bound anyway because there is no amplification: each envelope is paid for by one entity cell — its own mailbox, watch registration and registry entry — that had to exist first and was admitted through the region’s routing bound on the way in. A peer that can queue N of these has already made the node hold N actor cells, so the lane is a strictly smaller copy of a cost accepted upstream; and the envelope is enqueued as the cell goes away, replacing something larger than itself, so the peak falls rather than rises at the moment the traffic appears.

A source that broke that ratio — one cheap remote action queueing many envelopes, or envelopes tied to no resource the node already gates — is what would need an answer, and the answer would be per-kind coalescing (collapse duplicate suspend / resume for the same child) rather than a cap. Dropping is the one thing this lane may not do.

Envelope<unknown>

void

DroppingMailbox.enqueueSystem


hasMessages(): boolean

Defined in: src/internal/Mailbox.ts:566

boolean

DroppingMailbox.hasMessages


hasSystemMessages(): boolean

Defined in: src/internal/Mailbox.ts:570

boolean

DroppingMailbox.hasSystemMessages


hasUserMessages(): boolean

Defined in: src/internal/Mailbox.ts:569

boolean

DroppingMailbox.hasUserMessages


observeDrops(observer): void

Defined in: src/mailbox/DroppingMailbox.ts:84

See DropReportingMailbox.observeDrops — additive.

MailboxDropObserver<T>

void

DroppingMailbox.observeDrops


prependUser(envelopes): void

Defined in: src/mailbox/BoundedMailbox.ts:152

A replay meets the bound, the same way an arrival does (#772).

Until this override existed the base prependUser wrote straight to the queue, so unstashAll() unshifted a whole stash — up to DEFAULT_STASH_CAPACITY envelopes — past the capacity check, past the switch above and past the drop accounting. A reject mailbox never threw, a drop-head / drop-new mailbox never dropped, and droppedCount / actor_mailbox_dropped_total under-reported by exactly the batch. The advertised memory ceiling was not one: capacity: 10 could become 1034.

Which end sheds. The policy is unchanged, only the geometry is: an arrival lands at the tail and drop-head makes room at the head, a replay lands at the head and so makes room at the tail. Both read as “admit the arrival, evict a queued message from the far end”, and the other choice is indefensible here — evicting the head under a prepend would discard the messages the replay just put back, which is not a bound but a way of making unstashAll() a no-op.

Which reason is reported. drop-head when a queued message was evicted, drop-new when the arrival itself was refused — the closed two-value vocabulary of MailboxDropReason, applied by what actually happened rather than by which policy is configured. So a drop-head mailbox does report drop-new for the tail of a batch bigger than its capacity: once the queue holds nothing droppable, there is no room to make and the arrival is what goes. That is where this diverges from enqueue, which admits and overshoots by one in the same situation — one envelope over a bound is the arrival rate, a whole stash over it is the defect.

What reject does. It throws MailboxFullError, and it throws before admitting anything: all-or-nothing, so a caller that catches it knows the batch is entirely still its own. The throw lands on the actor’s own stack, inside its unstashAll(), rather than on a remote sender’s — and that is the closest thing to a sender a replay has. It is also not a message lost: ActorCell.unstashAll puts the batch back into the stash before the error travels on, so the envelopes stay parked and deadLetterStash still sees them if supervision then restarts or stops the actor. Choosing reject says “refuse, do not lose”, and refusing the replay of a stash that no longer fits is what that means here.

An Envelope.undroppable envelope is admitted whatever the policy says and is never counted, exactly as enqueueSignal admits one at the tail: a Terminated that round-tripped through a stash must not become droppable on the way back in (#729).

Envelope<T>[]

void

DroppingMailbox.prependUser


resume(): void

Defined in: src/internal/Mailbox.ts:576

void

DroppingMailbox.resume


suspend(): void

Defined in: src/internal/Mailbox.ts:575

void

DroppingMailbox.suspend