Things to know

Limits per connection and per run. A socket may have 1024 live queries and 64 running mutations/actions; queries time out after 30 s (--query-timeout), mutations after 5 s (--mutation-timeout); a client silent for 45 s (it answers pings) is disconnected. They bound the server's memory whatever a client does (docs/PROTOCOL.md, Limits).
Query results are cached while nothing they read changes. Every subscriber of a query shares one result; after the last one leaves it stays cached (up to 30 minutes, 64 MiB in all: --query-cache-ttl, --query-cache-size) and answers the next subscriber, HTTP query or ctx.run_query without running, until a write touches what it read. Repeated HTTP queries are cached too. A query that reads ctx.now() is never cached. The Overview page and /metrics show the hit rate.
Optimistic updates are replayed on top of every server change until the client has seen the mutation's writes (or it failed). Create ids and timestamps outside the update closure (once per call), not inside it:
ts
```
const id = `optimistic:${crypto.randomUUID()}` as Id<"messages">, now = Date.now();
send.withOptimisticUpdate((store, args) => insertAtTop({ …, item: { _id: id, _creationTime: now, …args } }));
```
Scheduled functions (and crons) run without an identity, whoever scheduled them: pass the user as an argument to an internal function, e.g. ctx.scheduler.run_after::<deliver_reminder>(delay, (channel, user, text)).
ctx.now() in a query is the snapshot's time, not a reactive clock. A query reruns when its data changes, not when time passes. For expiry, store the state and flip it with a scheduled mutation (run_at); format "5 minutes ago" in the client.
A query's result must depend only on its arguments and the data it reads. In --dev and TestApp, rxferrix reruns each query in the same snapshot now and then and logs a warning (also sent as a log line) if the result differs: SystemTime::now(), random numbers or HashMap order. Use ctx.now(), and a BTreeMap or a sort.
Invalidation is coarse: any write to a document reruns the queries that read it, even when nothing changed. Skip no-op writes, keep fast-changing fields (presence, typing, counters) in small tables of their own, read narrow ranges (eq prefixes, _creationTime bounds, take(n)), and count with .count().
Tests (rxferrix::testing::TestApp, in dev mode): t.create_user("ann@example.com", Some("password")) makes a user, t.as_user(id).query::<list>(args) calls as them, t.advance_time(d) runs due jobs.

Things to know ​

Things to know