MongoTimerStore

The TimerStore of one bounded context on MongoDB: a collection per aggregate type, wow_timer__<context>__<aggregate>, named by the rule of the log's collections, created on the type's first use and recorded in the registry wow_timer_aggregate. One document per timer, keyed by a unique index on its partition, its owner's tenant and id (owner: {tenant_id, id}, its context and aggregate the collection's) and key.

A timer is in its owner's partition, among the partitions of the context's log (Timer.partition). On a sharded cluster each collection is sharded by {partition: 1}, one chunk per partition, spread over the shards while it is new (as the log is): every filter by owner carries the owner's partition, so the reads and writes by owner and the due read of a node's partitions reach the shards of those partitions alone, and the key stays unique across the shards. A retry's read of a target's parked timers (parked) asks every shard.

Every write is one single-document operation, but for cancelAll and settle. A put is a conditional upsert (see conditionalUpsert) whose filter admits only another timer at the same or a later version; a cancel is a conditional update; a settle is one unordered bulk write of conditional updates per collection, each on the document as it was loaded. The due read is one aggregation over the registry: the count of the context's collections, then, through $unionWith, each collection's first due timers over its partial index on (partition, due_at, timer_id) of the PENDING and FIRING (claimed) timers, merged by due time. A count other than the collections read tells the store to read the registry again, so a collection another node created is read at once. The body is kept as the command's JSON text, the rule as a subdocument ({type, ...}, as Wow's JSON writes it). Format 5 is the layout per aggregate type; the timers are a view rebuilt from the log, so a collection of an earlier format is refused, dropped and rebuilt.

Calls run on dispatcher, virtual threads by default.