콘텐츠로 이동
한국어

Shard

이 콘텐츠는 아직 번역되지 않았습니다.

Defined in: src/cluster/sharding/Shard.ts:63

One shard of a sharded type — child of its ShardRegion, parent of the entities whose id hashes into it. Its entire job is the entity lifecycle: spawn, watch, stop, and buffer traffic for an entity that is on its way out.

Routing, buffering and coordinator bookkeeping stay in the region; the policy decisions about when to passivate stay there too (the region sees every message on this node, a single shard only sees its own slice). The shard just executes PassivateEntity. Keeping it that thin is what lets maxEntities go on meaning “per node” rather than silently becoming “per shard”.

Being a real actor is the point: a shard now has a path (/system/cluster/sharding/region-<type>/shard-<n>), so it is addressable from anywhere in the cluster, it shows up in the actor tree, and handoff is simply “stop the shard” — the runtime terminates the entities underneath and only then reports back.

new Shard(config): Shard

Defined in: src/cluster/sharding/Shard.ts:79

ShardConfig

Shard

Actor.constructor

readonly config: ShardConfig

Defined in: src/cluster/sharding/Shard.ts:79

get entityCount(): number

Defined in: src/cluster/sharding/Shard.ts:99

Number of live entities — the region mirrors this for its stats replies.

number

displayName(): string

Defined in: src/Actor.ts:192

Human-readable name for this actor in log lines and in the DevTools actor tree (#891). Defaults to the full path — which is already the log source, so an actor that doesn’t override this logs exactly as it did before.

override displayName(): string { return `User(${this.entityId})`; }

Purely cosmetic. The path stays the identity everywhere that routes, correlates or aggregates — metric labels, tracing attributes, dead letters, ActorRef.toString(), every wire identifier — so a display name is free to be ambiguous, unstable, or shared between actors.

Resolved on every record, not captured once. Two consequences: keep it cheap and side-effect free, and expect it to be called before preStart (hence the optional chain — the context is attached after construction). In exchange a name may be derived from state, and it updates when that state does. Throwing, or returning anything but a non-empty string, falls back to the path and warns once: a naming hook must not be able to take a log line down with it.

ActorOptions.withDisplayName(...) outranks this, for the same reason withSupervisorStrategy(...) outranks supervisorStrategy — the spawn site is the more specific statement. It has to: every Behaviors actor is a TypedActor that inherits this default, so a method that won would silently swallow the spawn-site value for exactly the actors that have no subclass to override. For a name that only becomes known at runtime, this.context.setDisplayName(...) outranks both.

string

Actor.displayName


onReceive(message): void

Defined in: src/cluster/sharding/Shard.ts:86

Arm order is deliberate: every message to an entity lands here, and the two instanceOf checks are the expensive ones — putting the envelope first keeps the hot path a single string comparison.

ShardInbox

void

Actor.onReceive


postRestart(_reason): void | Promise<void>

Defined in: src/Actor.ts:152

Called on the fresh instance after a restart. Default: call preStart().

Error

void | Promise<void>

Actor.postRestart


postStop(): void | Promise<void>

Defined in: src/Actor.ts:103

Called after the actor has been terminated. Children are already stopped.

void | Promise<void>

Actor.postStop


preRestart(_reason, _message?): void | Promise<void>

Defined in: src/Actor.ts:119

Called before a restart, on the instance about to be thrown away. The default calls postStop() and nothing else.

Override to release what the instance holds outside itself — a file handle, an open socket, a broker connection — or to do something other than drop the message that failed.

Stopping this actor’s children is not done here: the framework tears them down after this hook returns and waits for them before building the replacement, because postRestart re-runs preStart and a named child needs its name back. To keep the children instead, see Actor.stopChildrenOnRestart.

Error

ShardInbox

void | Promise<void>

Actor.preRestart


preStart(): void | Promise<void>

Defined in: src/Actor.ts:100

Called after construction and before the first message is processed.

void | Promise<void>

Actor.preStart


stopChildrenOnRestart(): boolean

Defined in: src/Actor.ts:149

Whether a restart tears this actor’s children down before rebuilding it. Default: true.

A restart replaces the Actor instance while the cell — and therefore the child map — survives. Keeping the children was the old behaviour and it made an ordinary pattern impossible: postRestart re-runs preStart, so an actor that spawns a named child there hit Child name … is not unique on its first restart and never recovered (#634).

Override to false when the children are expensive to rebuild, hold state the parent cannot restore, or are supervised independently — a connection pool, say. They then outlive the restart exactly as before, and it is on you to make preStart idempotent — by adopting the survivor from the cell, this.child = this.context.child('name').toNullable() ?? this.context.spawn(Child, 'name'), or with context.spawnAnonymous. An instance field cannot do it: preStart runs on a fresh instance after every restart, so this.child ??= … is always unset and re-spawns into the name the surviving child still holds, which fails the spawn and restarts the actor again.

This is a separate hook rather than a preRestart override because the teardown has to be awaited: the new instance cannot be built until the old children are actually gone, and preRestart has no way to tell the cell that it started something worth waiting for.

boolean

Actor.stopChildrenOnRestart


supervisorStrategy(): SupervisorStrategy

Defined in: src/Actor.ts:160

Supervisor strategy for this actor’s children. Defaults to restart, up to 10 times per minute, then stop.

SupervisorStrategy

Actor.supervisorStrategy