Storage and retention¶
RF Sentry stores large evidence in browser-local IndexedDB and OPFS under the current HTTPS origin. Use a stable canonical origin; browsers intentionally do not share one origin's archive with another hostname.
What lives where¶
| Content | Primary storage | Resident working set |
|---|---|---|
| Raw DVR IQ | OPFS warm segments + hot RAM | configured hot tier and replay cache |
| Decoder events | bounded worker/event store | visible tail and indexes |
| Aviation observations | IndexedDB indexed records | current snapshots and paged ranges |
| FCC corpus | SQLite WASM in OPFS | indexed cache or explicit eager index |
| SATCOM originals/native derivatives | OPFS | visible/nearby preview/native frames |
| Preferences | IndexedDB/localStorage as appropriate | compact state |
IQ DVR controls¶
Pause/Play, seek/position controls and Go Live review captured raw IQ while
the physical receiver continues receiving. Space controls transport;
tap/hold W and S seeks. Historical VFO changes remain inside the captured
RF span, including the entire selected passband. Hardware controls stay
read-only during historical review. Event-table Freeze view is different:
it freezes one review surface while event ingestion continues.
| Setting | Default | Meaning |
|---|---|---|
| Hot RAM Minimum | 60 seconds | recent raw-IQ duration held in RAM |
| Hot RAM Target | 1 GB (decimal) | nominal raw-IQ payload target, not a hard RAM ceiling |
| Total retention | 6 hours | requested age across RAM and OPFS |
| Warm storage budget | 100 GB (decimal) | browser-local IQ plus waterfall payload budget; quota can be lower |
| Pressure policy | Protect history | prioritize the interval being investigated |
The minimum duration wins when its IQ needs more memory than the target; the UI reports that conflict. Reducing the target requests a rebalance, not data quality reduction. OPFS faults and history-loading states are explicit. Storage is on the workstation, not on the Pi. Per-lane preferences are saved locally; changing origin does not carry those preferences or history with it.
Under Protect history, pressure can stop recording newer live history to keep the selected interval. Protect live instead returns you to Live. Go Live normally preserves history; it may clear/restart the timeline only after newer live history was deliberately sacrificed. Neither policy pauses the physical receiver. Returning to Live is explicit, not an automatic claim that accelerated review is current reception.
Waterfall quality and memory¶
Visual history defaults to native emitted Float32 rows. Optional peak-preserving 1/2, 1/4 or 1/8 resolution and 16/8-bit dB encoding are explicit choices for future visual-cache rows; they neither reduce raw IQ nor alter the selected scientific FFT. A preference change does not discard or synchronously repack existing rows.
The DVR panel separates raw-IQ and waterfall-row payload. The nominal target does not include every JavaScript, worker or GPU allocation, so combined payload can exceed it. The raw-IQ tier rebalances toward 90% after crossing or reducing its target, unless the minimum duration itself requires more memory.
Original timeshift acceptance and the newer tiered-storage gate are distinct; see DVR qualification. Do not treat rolling retention as a permanent recording/export guarantee.
Client memory readouts¶
The header's App RAM ≈ reports retained hot IQ, the warm replay cache, waterfall-row payload, FCC SQLite WASM linear memory plus a sampled hot-index estimate, decoder/channelizer/media allocations and DSP WASM linear memory. GPU ≥ separately reports owned WebGPU/WebGL spectrum and FFT buffer/texture descriptors; it is a lower bound and is not added to the RAM figure.
Worker allocations are sampled at low cadence without a whole-origin memory or garbage-collection diagnostic during reception. Browser/worker overhead, other JavaScript objects, MapLibre basemap textures, GPU compositing surfaces and transient FCC staging are unmeasured. These are app-accounted allocations, not total process memory, system RAM or physical VRAM usage.
Chrome's legacy JavaScript-heap used/limit pair is a separate, coherence-checked advisory with warning/critical levels at 80/90% after sustained high readings. Inconsistent readings do not produce a pressure percentage. The estimate can miss workers or include other pages; it is not free workstation RAM or a whole-browser memory measurement. A Hot RAM Target/minimum conflict is a DVR policy advisory, not proof that the workstation has exhausted memory.
FCC mode¶
Indexed is the default unless the operator has explicitly saved Eager. Eager warms a larger SQLite page cache for deeper regulatory work with visible progress and cancellation. Both modes keep compact license summaries and callsign indexes in RAM; changing modes does not erase records or switch to another corpus. Legacy implicit Auto preferences resolve to Indexed.
The status distinguishes requested mode, effective mode and loading progress. Eager preparation scans in bounded batches and can be cancelled; the existing indexed runtime supplies compact enrichment during warming. Peak-memory admission can decline Eager with an explicit reason rather than starting a second full worker. Cancellation/failure falls back to Indexed.
Switching back to Indexed closes and terminates the owner before starting its replacement, because grown WASM linear memory cannot shrink. Cancellation or a failed Eager transition also recycles to release that growth. Expect a temporary FCC-readiness/reload state and interrupted pending FCC queries during recycling; the persisted corpus and separately learned RF history remain intact, and no receiver restart is needed. An Indexed transition must not secretly retain the old eager worker alongside its replacement.
Configured SQLite cache capacity is not the same as measured resident memory. Browser process memory also includes workers, decoded surfaces, GPU resources, WASM heaps, and internal caches.
Canonical-origin warnings matter: another hostname has separate IndexedDB, OPFS and preferences even when it reaches the same appliance. Import/update failure keeps the prior valid dataset where available; unavailable enrichment must not prevent station learning. Source attribution, generation status and lossless-import details are in FCC datasets.
Aviation history¶
The default browsing horizon is 24 hours, but detailed records are paged. Old ranges remain searchable and replayable from persistence; inactive page caches age out even if the decoder is stopped. Deleting a cache must not delete raw observations.
SATCOM imagery¶
Native products remain in OPFS. Gallery thumbnails and animation frames use bounded derived caches. Native playback may reserve substantial memory under an operator budget, but decoded surfaces, transfers, and GPU copies all count. Jobs share prepared frames by identity and release them when no consumer needs them.
Browser data controls¶
Archive deletion is destructive and should identify exact scope. Clearing site data removes origin-scoped IndexedDB, OPFS, and preferences; it is not a normal migration strategy. Export important evidence before clearing.
Closing the tab/browser ends browser-owned acquisition and processing. It does not move work to the Pi in the background.
Optional rfSentry Server decoded-record archive¶
The separately installed Server runs on another machine. Its React administration UI is served by C++; the installed package needs no Node runtime. Bind an explicit archive volume, pair the appliance or register an authorized narrow credential, then enable each desired stream. Pairing and upgrades preserve inactive policies. The Pi retains decoded records only in RAM and owns acquisition/decoding; Server storage faults and backlog never control tuning or create another SDR connection.
The Phase 6 archive stores exact canonical envelopes, configuration/boundary dependencies, timing and validation evidence. GOES HRIT payloads remain the original 892 post-RS bytes. ACKs follow segment synchronization and durable local SQLite index/resume commits. Ring expiry and source restarts appear as explicit gaps; continuing from available history requires a saved operator decision.
Age/byte retention is explicitly enabled per stream. Age uses first Server receipt UTC; deleted identities remain recorded and retries do not resurrect them. Root changes create an independent archive namespace and never inherit another root's cursor. Both processing intentions share this archiving stage; product reconstruction remains unavailable until Phase 8. Browser source integration and peer hydration are not part of Phase 6.
See the repository's components/rfsentry-server/README.md for setup and controls,
and components/rfsentry-server/ARCHIVE.md for storage format, filesystem
assumptions, recovery, retention and bounded coverage/readback APIs.