Availability and resource limits
This page lists what is expensive in iya-sts, how the service stops overuse from taking it down, and how it avoids making a caller wait past its own timeout.
What is expensive
| Feature | Why | Defence | Setting or limit |
|---|---|---|---|
| Password hashing (scrypt) | Costs memory and CPU on purpose | Runs on libuv’s thread pool, off the event loop. Attempts are rate limited per identity and per address, and a growing wait follows each refused password | security.rateLimitPerIdentity (5), security.rateLimitPerAddress (20), security.rateLimitWindowS (60). The compose stack sets UV_THREADPOOL_SIZE to 32 |
| Post-quantum signing and key generation | Slow algorithms | Run natively on the same thread pool | Off the event loop |
| Protocol request handling | Every request runs on one of a few threads | Worker threads run the protocol stack, so the listener is not blocked. Each thread has a heap limit, and one that reaches it is replaced | workers.requestCount, workers.surfaceCount, workers.heapLimitMb |
| SCIM bulk and Shared Signals receive | Large writes | A dedicated lane, so a bulk load cannot fill every worker. Past the queue, the answer is 503 with Retry-After |
workers.batch, workers.batchWorkerShare (50%), workers.batchConcurrency (8), workers.batchQueueLimit (5,000), workers.batchQueueTimeoutS (60), scim.bulkMaxOperations (100), scim.bulkMaxPayloadSize (1 MiB) |
| Directory searches | A broad filter can read many rows | Paged in a windowed process, stopping at the first limit reached. The partial answer is marked incomplete | ldap.sizeLimit (500), ldap.timeLimit (30 s) |
| Reports | Can read a whole realm | Bounded synchronously, and queued above that. Limited per realm and per person | report.syncMaxRows (500), report.syncMaxBytes (1 MB), report.syncMaxMs (10 s), report.asyncMaxRows (100,000), report.asyncMaxBytes (32 MB), report.asyncMaxMs (120 s), report.maxConcurrentPerRealm (2), report.maxConcurrentPerSubject (1) |
| Risk dataset import | Files up to 2 GiB | Streamed and imported in the background. The upload answers as soon as the file is stored | risk.uploadMaxBytes, risk.expandedMaxBytes, risk.expansionMaxRatio, risk.importStallMinutes (15) |
| Periodic jobs | Every realm’s hourly job falls due together | At most a fixed number run at once, each with a time limit | scheduler.maxConcurrentRuns (0: two runs, or one fewer than the request worker threads when that is more), scheduler.runTimeoutS (600) |
| Outbound requests | A partner may be slow or hostile | A timeout and a response cap. A fetch not needed for the answer is never awaited | federation.outboundTimeoutMs, federation.maxResponseBytes |
| Compressed and XML messages | Can expand without bound | Inflated up to a fixed limit, parsed with no document type | Fixed |
| Request bodies | Memory | 5 MB at most, unless a field states less | Fixed |
| Requests from one address | Any endpoint, repeated | A ceiling per address per minute in each process. Past it, the answer is 429 with Retry-After: 60, and one warning per address per minute is logged (STS-CORE-0172). /healthcheck is not counted |
security.requestsPerAddressPerMinute (6000 in product mode, 0, no ceiling, in development) |
| Connections | File descriptors and memory | Limited per listener | http.maxConnections, http.maxRequestsPerSocket, http.headersTimeoutS, http.keepAliveTimeoutS |
| Sessions and in-flight sign-ins | Memory and store rows | Capped. At the cap a new row is refused, and nothing is ended to make room for somebody else | authn.maxSessions (100,000), authn.maxSessionsPerSubject (32), oidcRp.maxFlows (200) |
| Shared Signals push | A slow receiver | One queue per receiver. Past it, an event is dead-lettered | ssf.pushBacklog (2,000) |
| The audit log | Grows without bound | A ring. The oldest rows drop, and the page shows how many | audit.maxEvents (5000) |
Answering before the caller gives up
The service answers quickly or tells the caller to come back. It does not hold a connection for slow work.
- Queued work. A large report is a job: the caller gets an id and polls for the result.
- Answer first, finish later. A risk dataset answers when stored, and
POST /ssf/verifyanswers204when queued. - Polling protocols. ACME orders, the device flow and CIBA return at once and the client polls.
- Refuse rather than wait. A full batch lane answers
503withRetry-Afterwithinworkers.batchQueueTimeoutS. - A search ends with a partial answer at its time limit.
The HTTP server’s own request timeout is five minutes. The whole of a very large
upload has to arrive in that time. For a very large risk dataset, use
risk.datasetsDirectory.
What you must add
The service limits credential attempts and the costly features above. It also
puts a ceiling on all requests from one address:
security.requestsPerAddressPerMinute, 6000 a minute in product mode and off in
development. Behind a proxy, set global.trustedProxies so the address is read
correctly. The ceiling is counted per process, so a cluster of N nodes allows N
times as much. There is no per-client limit and no limit on how many requests
one client can run at once. Put a rate limit sized for your users in front of
the service, in your load balancer or gateway, and size the thread pool and the
workers for your load.