Admission Control

SIPI admits requests through a cost-based two-partition pool that protects viewer tile traffic and the RAM envelope from expensive full-image decodes (overwhelmingly distributed crawler bots). Each request is classified into a Tile or Full partition; tiles get a guaranteed thread floor and burst into idle capacity, full-image downloads are hard-capped in both threads and decode memory and shed when their share is exhausted.

This page covers the thread partitions and the overall model. The memory half — the full partition's decode-RAM cap — is detailed in Memory Budget.

The model

Everything derives from two hard caps (RAM + CPU threads), two ratios, and a mode:

• Threads. tile_min = round(nthreads × tiles_thread_ratio) is guaranteed to tiles; full_max = nthreads − tile_min hard-caps concurrent full decodes. A tile takes a global permit only and may burst up to nthreads when the full partition is idle. A full takes its full-partition permit first, then the global permit, so at most full_max fulls ever contend for global permits and ≥ tile_min are always free for tiles.
• Memory. full_mem = memory_limit × (1 − tiles_memory_ratio) caps the full partition's decode bytes; tiles bypass the budget. See Memory Budget.
• Mode. Two tiers. The basic tier (the global CPU/thread concurrency cap) is always enforced. basic (default) enforces only that tier and shadow-counts what the advanced tier (the full thread cap plus the memory budget) would shed; advanced also enforces the advanced tier and sheds with 503/413.

Tile priority is bounded, not absolute: tokio's semaphore is FIFO, so a full already parked in the global queue does not yield to a later tile. The tile floor (≥ tile_min global permits kept free of fulls) is the mechanism that preserves tile headroom — an explicit design choice, not automatic preemption.

What is a tile vs a full request

Request	Partition
Viewer tiles, small explicit {w},{h} / !{w},{h} sizes	Tile
info.json, knora.json (metadata, no decode)	Tile
/{id}/file (raw byte stream, no decode)	Tile
Lua routes (global permit only; any decode they trigger is still memory-budgeted)	Tile
Large explicit sizes, /full/max/, percentages (estimated peak ≥ large_decode_threshold_bytes)	Full

Classification is coarse in the shell (from the IIIF URL params, before any decode) and precise in the engine (from estimate_peak_memory). The two are tracked and their disagreement is observable, so residual drift can be tuned.

Configuration

Env var	CLI flag	Default	Description
SIPI_NTHREADS	--nthreads	0 (auto)	Worker threads (0 = auto-detect)
SIPI_TILES_THREAD_RATIO	--tiles-thread-ratio	0.5	Fraction of workers guaranteed to tiles (0..1)
SIPI_MEMORY_LIMIT	--memory-limit	0 (auto)	Total RAM envelope (0 = auto-detect)
SIPI_TILES_MEMORY_RATIO	--tiles-memory-ratio	0.25	Fraction of the envelope reserved for tiles + the non-decode floor
SIPI_ADMISSION_MODE	--admission-mode	basic	basic (enforce basic tier only) or advanced (also enforce the advanced tier)
SIPI_LARGE_DECODE_THRESHOLD_BYTES	--large-decode-threshold-bytes	33554432 (32 MiB)	Estimated peak at/above which a decode is a full-partition decode

ops-deploy renders DSP_IIIF_MEMORY_LIMIT → SIPI_MEMORY_LIMIT and DSP_IIIF_ADMISSION_MODE → SIPI_ADMISSION_MODE. An unrecognized admission_mode (e.g. a stale off or a legacy monitor/enforce) is not a startup error; it silently falls back to the basic default.

Basic → advanced workflow

1. Ship in basic (default). No ops-deploy change needed: the defaults live in the binary, so the fingerprint and shadow counters appear on Grafana as soon as the new binary runs. basic enforces only the basic tier — the advanced tier is observe-only, so it changes no behavior (the full thread cap does not reject; it shadow-counts).
2. Observe (1–2 weeks):
3. sipi_admission_permits_in_use / sipi_admission_permits_total and sipi_admission_full_in_use — global and full-partition saturation.
4. sipi_admission_tile_waiting / sipi_admission_full_waiting, sipi_admission_tile_shed_total / sipi_admission_full_shed_total — per-partition queue pressure and 503s.
5. sipi_admission_full_shadow_rejected_total and the full-partition memory shadow counters (see Memory Budget) show what advanced would shed — enough to size full_max/full_mem.
6. sipi_admission_mode, sipi_admission_tile_min_threads, sipi_admission_full_max_threads, the ratios, sipi_admission_memory_limit_bytes, sipi_admission_large_decode_threshold_bytes — the config fingerprint.
7. Tune the ratios / memory_limit if the shadow counters fire on legitimate traffic.
8. Switch to advanced: set DSP_IIIF_ADMISSION_MODE=advanced and redeploy.

Rejections

• 503 Service Unavailable + Retry-After — the pool (threads) or the full memory budget is currently saturated; retry may succeed.
• 413 Payload Too Large (no Retry-After) — a single request's estimate alone exceeds the full-partition memory budget; it can never succeed.

Tiles are never shed for full-partition pressure — they only shed when no global permit is genuinely free (a tile burst beyond nthreads).

Metrics

All admission metrics share the sipi_admission_* namespace (rendered from the OTLP sipi.admission.* instruments; Prometheus appends _total to counters).

Gauges (point-in-time occupancy + sizing):

• permits_in_use / permits_total — global pool saturation.
• full_in_use — full sub-pool permits held (against full_max_threads).
• tile_waiting / full_waiting — requests parked per partition. Tiles wait only behind other tiles (exempt from the full queue-depth shed).

Counters (monotonic):

• tile_shed_total / full_shed_total — 503 sheds per partition.
• full_shadow_rejected_total — basic-only: fulls the advanced-tier cap would have rejected (zero in advanced, where full_shed_total counts the real rejections). The signal that sizes full_max before the switch.
• classifier_disagreement_total — serves where the shell's pre-dispatch tile/full verdict differed from the engine's precise post-decode verdict. A low, flat value confirms the pixel-proxy heuristic tracks the engine; a rising value means the bytes-per-pixel proxy needs revisiting.

Config fingerprint (gauges, observable with no ops-deploy change): mode, tile_min_threads, full_max_threads, tiles_thread_ratio, tiles_memory_ratio, memory_limit_bytes, large_decode_threshold_bytes.

The full-partition memory metrics (decode_memory_*, including the 413/too_large counters) are documented in Memory Budget.

Temporality. The sipi_admission_* counters are cumulative (monotonic) OTLP sums that live for the whole process and reset only on restart, so rate() / increase() read them correctly. (The max_over_time() idiom some SIPI dashboards use is for windowed extremes over gauges, a different query pattern — it does not apply to these counters.)