--- name: worker-threading description: When map engines move work off the main or render thread (Web Workers, OffscreenCanvas, WASM threads, native thread pools), what they send across, how they size, start, cancel and recover workers, and what they reverted, distilled for a map-from-scratch renderer. Load it before adding a Worker, worker_threads, SharedArrayBuffer or a thread pool to a renderer or to bake.ts, when data must be decoded, triangulated or laid out at run time, when a bench shows main-thread jank during loading, or when a worker path misbehaves (a blank map on GitHub Pages, memory growth, stale or lost results, output that changes with the thread count). --- # Worker threading: what to move off the main thread, and what it costs A worker keeps the frame loop free while data is decoded, triangulated or laid out, but every engine that adopted workers then paid for them: transfer and message round trips, worker start-up and duplicated bundle code, cancellation and stale-result bugs, browser-specific leaks, and, with SharedArrayBuffer, a requirement on the host's HTTP headers. The 37 engines behind this skill also show the opposite case: work that moved back to the main thread, worker pools cut to one, and decode workers reverted because the network, not the CPU, was the bottleneck. The map-from-scratch rule is to do the work at bake time, so the renderer receives data laid out for exactly what it draws, and no example so far uses a worker [2]. Load this skill when that stops being enough. Per-engine detail, with more numbers, is in [references/engines.md](references/engines.md). Neighbouring skills: `tile-cache-and-loading` (scheduling, cancellation and caches), `data-range-requests` (Range reads), `label-placement` (where collision runs), `polygon-triangulation`, `raster-tiles` (image decoding) and `text-rendering` (font workers), `startup-and-bundle-size` (worker bundle bytes) and `mobile-and-power` (workers killed on iOS). ## First: is there work left for a worker? Every engine record that recommends workers adds "unless the data is baked". Check these before writing one: - **The data is baked and loaded whole.** The renderer only uploads buffers and draws. A worker adds start-up and a copy and removes nothing. MapLibre GL JS itself defaults to one worker on every browser but Safari, after A/B testing, where it used to start up to six [5]. deck.gl deferred its RFC to build layer attributes in a worker: review found updates could be slower because of the message exchange, and it did not work for composite layers such as GeoJsonLayer [6]. - **Worker start-up can exceed the work.** In ortho-earth the first decode call was dominated by worker module evaluation, 1,538 ms, against 26 + 20 ms to parse on the main thread [7]. Cesium's move of WASM decoders into workers cost tens of milliseconds of cold start on first use [8]. - **Measure where the time goes first.** ortho-earth reverted three COPC reader workers because reading waited on the network (settling varied 2.4-4.0 s with the network; decode of 43 nodes took 1.4 s in total) [9]. Its GPU-capable stages were 3-6% of a tile conversion, so moving them helped little [10]. QGIS found that PNG encoding, not rendering, dominated tile export (268 s to 21 s once encoding moved to a pool) [11]. - **Image decoding needs no worker of your own.** `createImageBitmap` lets the browser decode off the main thread. RustGIS's 20-30 ms wasm decode per tile vanished from browser profiles when it switched [12]. Cesium 1.56 made ImageBitmap the default; 1.56.1, the next day, went back to Image and made ImageBitmap opt-in, noting that an ImageBitmap cannot be flipped during texture upload [13]. See `raster-tiles`. Run-time work that does justify a worker: decoding and triangulating large streamed data (MVT, MLT, GeoParquet, LAZ) [14][15], per-pixel raster operations [16], text shaping with fonts unknown at bake time [17], and live data that changes faster than the main thread can re-index it [18]. ## What goes to the worker, and what stays - **Do the heavy one-shot work in the thread that already parses the data.** deck.gl moved earcut into the loaders.gl worker that parses MVT, rather than a second deck.gl worker, which would have moved the data across threads twice. Binary MVT parsing was 1.9-2.4x faster in tiles/s, and a full-viewport load of 988K vertices fell from 6,242 ms to 2,713 ms; a 12K vertex set gained only 13% [14]. The RFC warns that for other loaders the transfer may cost more than the work saved [14]. - **The main thread uploads and draws.** MapLibre GL JS fetches, decodes, builds buckets, indexes features and lays out symbols in workers, and uploads vertex buffers lazily on the main thread [19]. - **Global decisions come back to the main thread.** MapLibre GL JS and Native moved label collision out of per-tile worker jobs into one viewport-wide index on the main or render thread, after a series of perspective workarounds in the worker [20][21]. Tangram JS split it: a per-tile pass in the worker, a global pass on the main thread fed only a box and a buffer range per label [22]. See `label-placement`. - **Small per-frame updates stay where the frame is drawn.** MapLibre Native added a per-source `synchronousUpdate` so a location marker that changes every frame does not lag a frame behind, and limits it to small data [23]. OpenLayers' worker-built point buffers lag the view, so each result carries the transform it was built with and hit detection is skipped while out of sync [24]. - **If you draw in a worker (OffscreenCanvas), draw everything there.** ortho-earth's overlays drawn on the main thread with the current camera ran 1-2 frames ahead of the worker-rendered globe; drawing them in the worker's requestAnimationFrame removed the lead [25]. ## What crosses the boundary - **Flat typed arrays, transferred.** MapLibre GL JS stores collision boxes in a StructArray: about 500 KB less memory per tile, about 60% faster creation, and transferable [26]. ortho-earth sends attributes as one typed-array table with dictionaries; worker start-up for 1,000,000 features fell from 16 s to 1.6 s [10]. deck.gl calls the many small JS arrays of GeoJSON-shaped output "very slow" [27]. MLT's wasm decoder hands over whole columns: 524,482 vertices decoded in about 3.8 ms [15]. - **A transferred buffer is gone from the sender.** Copy deliberately what both sides need; MapLibre copies the raw tile data for that reason [28]. ortho-earth stored detached, empty buffers in IndexedDB until it awaited the write before transferring [29]. - **Send only what the receiver needs, and only deltas.** Tangram posts back a whitelisted slice of the tile; feature geometry stays in the worker [30]. MapLibre stopped sending the whole GeoJSON back after each update and applies the same diff on the main thread: 33.3 ms to 16.7 ms on its bench, at the price of two copies kept in sync [31]. - **Post setup data to the worker first, then process it locally.** MapLibre sends style layers to the worker before processing them on the main thread, so both work in parallel [32]. - **Do not stringify messages.** Tangram JSON-stringified worker messages on the strength of a 2016 benchmark and removed it because the gain was inconclusive on current browsers [33]. - **A transfer format borrowed from a tile spec imposes its value model.** MapLibre serialises worker-built GeoJSON tiles as MVT; undefined property values broke the encoder, and nested objects cannot be carried at all [34]. Your own format between your own threads can match your data. - **WASM memory views go stale.** navara re-takes `memory.buffer` views after any call that can grow WASM memory instead of holding one [35]. - **Test long sessions on Safari.** Mapbox GL disabled Transferables on Safari for unbounded memory growth; MapLibre re-enabled them after Safari fixed the bug [36]. ## SharedArrayBuffer: never required SharedArrayBuffer needs cross-origin isolation (COOP and COEP headers on the page). maplibre-rs's shared-memory web build does not work on Safari, hit memory corruption with a parallel worker in Firefox, and errors without the headers; a single-threaded path with transferred flatbuffers was added beside it [37]. ortho-earth uses SharedArrayBuffer only as an optimisation: without isolation it copies once, and a test runs seven pages with isolation off and asserts that the fallback was used [38]. In this repository the trap is concrete. The harness server sends COOP and COEP so that `performance.now()` has 5 us resolution [1], and `bin/mfs bench --site` serves the built Pages site through the same server [3], but GitHub Pages sends neither header [2]. A path that needs SharedArrayBuffer passes every bench and breaks for every visitor. If you use it, add a switch that forces the fallback and a verification check that fails unless the fallback ran. ## Pool size, start-up and lifetime - **Start with zero or one worker.** MapLibre went from half the cores to half capped at 6, then to 1 except on Safari (half the cores, at most 3), which is still the default [5]. ortho-earth keeps a persistent pool per kind with at most 2 lanes, shut down after 45 s idle, instead of a worker per call [7]. - **Create workers lazily.** OpenLayers creates its buffer worker on first use [39]. Cesium loads its 712 KB Draco decoder only inside a worker when Draco data appears; bundling it had raised gzipped Cesium.js by 18% [40]. - **Assume the browser can kill a worker.** iOS Safari kills workers without notice; MapLibre pings them on each return to the foreground and rebuilds the pool and the style if they stay silent, and lets the pool drain when the last map is removed [41]. - **Cap in-flight work by memory, not cores.** reearth-papers caps concurrent renders per isolate although rendering is synchronous, to bound resident tile data; unbounded concurrency produced RangeErrors [42]. ## Shipping the worker code - **Make the worker URL explicit or inline it.** MapLibre v6 requires `setWorkerUrl` because guessing the path broke under bundlers; an app that forgets it draws only the background [43]. Cesium embedded its workers in the single-file build to fix cross-origin loading [44]; Tangram builds the worker from a Blob URL, which needs a CSP allowance [45]. - **One worker entry, not one per role.** Cesium's workers fell from 2.6 MB to 225 KB gzipped once shared code went into shared chunks [46]. ortho-earth spawns one entry with the role in the Worker `name`; duplicated bytes fell from 2.74 MB to 177 KB [47]. - **Avoid nested workers.** In ortho-earth a worker spawned by a worker sometimes never ran (headless: 1 in 3 to 1 in 5 succeeded); the main thread now creates both and passes a MessagePort [47]. - **No CDN worker.** deck.gl loads its worker bundle from unpkg by default [48]; this repository allows no run-time external dependency [2]. Here: the harness server strips `.ts` and serves it as JavaScript [1], so `new Worker(new URL('./worker.ts', import.meta.url), {type: 'module'})` works under `bin/mfs bench`. `tools/build-site.ts` follows only `import` and `from` specifiers and rewrites `.ts` to `.js` [49], so on Pages that worker file is neither copied nor renamed. Either build the worker from a Blob of a string in the main module, or extend `build-site.ts` (with a test) to follow `new URL(…, import.meta.url)`. List the worker file in `chosen.code` [4]. ## Scheduling: cancel, coalesce, version - **A cancelled request must never call back, and must settle.** MapLibre nulls the callback before posting the cancel (unbounded memory growth in IE otherwise), rejects aborted promises with AbortError, and removes unloaded tiles from every worker; aborts "happen constantly during map movement" [50]. - **Latest wins.** MapLibre's GeoJSON worker and Native's geometry tile worker drop all but the newest pending job [18][51]. The first MapLibre version was reverted because dropped jobs never called back [18]. A later update queue chained every update and kept each one alive, so setData on every frame grew the heap until it ran out of memory [31]. - **Tag results with a generation.** tangram-es keeps old tiles visible and accepts a finished tile only if its source generation is newer. It first added "stale" and "updating" tile states ("updating" kept a stale tile already being rebuilt from being enqueued again), then replaced both with a per-tile source generation compared against the data source's [52]. - **Do not block a worker on a dependency.** tangram-es requeues a tile waiting for a raster instead of blocking the worker [53]. MapLibre batches glyph and icon requests to the main thread into 2 per tile [54]. - **Coalesce GPU rebuilds when results trickle in.** Once Cesium decoded splats in workers, tiles finished one per frame and each triggered a full re-upload; rebuilds are now debounced with a deadline [8]. - **Long jobs need fine-grained cancellation.** QGIS added exit points throughout label placement because cancelled renders kept burning threads [55]. ## Threads in bake.ts and native renderers Where the parallel work is a bake step (Node `worker_threads`) or a C renderer, the failures are shared state: - **Output must not depend on the thread count.** ortho-earth's converter writes byte-identical output at any worker count [10]; GRASS checks parallel modules against serial output [56]. Make it a verification check. - **No hidden shared state.** GRASS threads read each other's rows through one shared mask file descriptor (fixed by going serial with a mask) [57]; Marble wrapped a vendored decoder with static locals in a mutex [58]; a per-cell write of the same value to one shared flag stopped r.mapcalc from scaling [59]. Give each worker its own handle, buffers and seeded random stream; a shared atomic RNG was correct but 50-55% slower at 4-16 threads [60]. - **Scatter into per-worker buffers.** Unsynchronised `+=` into a shared grid lost 12-14% of the total in GRASS r.sim; the atomic fix cost 6% single-threaded [61]. - **Split work so no item is lost.** A hand-rounded chunk size skipped up to (threads - 1) / 2 items per pass in r.sim, and Marble left an unpainted strip when the height did not divide by the thread count; test counts not divisible by the worker count [62]. - **Band large rasters with halos, bounded by memory.** GRASS r.param.scale reached 4.6x on 8 threads with per-thread ring buffers and output identical to serial [63]. Encode in a pool and write from one thread (QGIS, above [11]). - **Default to all cores in a one-shot tool, pin inner tools to one thread when the caller already runs jobs in parallel** (GRASS) [64]. ## Tried and reverted | Change | What happened | Ref | | --- | --- | --- | | MapLibre: up to 6 workers | cut to 1 (Safari up to 3) after A/B testing | [5] | | MapLibre: per-tile collision in workers | moved to one main-thread index | [20] | | MapLibre: setData coalescing v1 | reverted (dropped callbacks), re-landed | [18] | | Mapbox GL: no Transferables on Safari | re-enabled after Safari fixed the leak | [36] | | Cesium: ImageBitmap by default (1.56) | back to Image, ImageBitmap opt-in, in 1.56.1 the next day | [13] | | Tangram: JSON-stringified messages | removed, no measurable gain | [33] | | Tangram: random tile-to-worker assignment | reverted to round robin; later each non-tiled source pinned to one worker | [65] | | navara: font worker merged into the general worker | reverted | [17] | | ortho-earth: three COPC reader workers | reverted, network-bound | [9] | | ortho-earth: nested workers | main thread creates both, MessagePort | [47] | | deck.gl: layer attributes built in a worker | RFC deferred | [6] | | GRASS: worker threads on by default (2008) | default set to zero in 2015 | [66] | ## Guidance for a renderer here 1. Do the work at bake time and ship typed arrays laid out for the draw calls. Then there is nothing for a worker to do; write "no worker" and the reason in `architecture_selection.not_doing`. 2. Add a worker only for measured run-time work: a long task on the main thread during load or interaction, not a guess. Record the before and after in `performance_knowledge`. 3. Use one module worker, created on first need, doing decode and triangulation in one place, returning transferred typed arrays and nothing else. Keep global decisions (label collision, picking state) and per-frame updates on the main thread. 4. Every request carries an id and a generation; cancel drops the callback and settles the promise; pending jobs coalesce to the latest. 5. Never require SharedArrayBuffer. Make the worker load on Pages (Blob or `build-site.ts` support), and run the page once without COOP/COEP. 6. The harness's `requests`, `transfer_kb` and `js_heap_mb` come from the page target only, and only page exceptions end a run [67]. Count the worker's fetches yourself (as with `range_requests` [4]), watch `browser_rss_mb` for its memory, and forward worker errors to `emit`. Whether CPU throttling reaches the worker is not verified here; check it before trusting throttled numbers. 7. In `bake.ts`, parallel output must be byte-identical to serial, checked by a test that runs both. ## Sources 1. `harness/lib/serve.ts` (this repository) 2. `README.md` and `skills/map-from-scratch/SKILL.md` (this repository) 3. `harness/cli.ts` (this repository) 4. `skills/map-from-scratch/references/spec-yaml.md` (this repository) 5. MapLibre GL JS worker count: https://github.com/maplibre/maplibre-gl-js/commit/58a61d625, https://github.com/maplibre/maplibre-gl-js/commit/ae427dc27, `ref-repos/maplibre__maplibre-gl-js/src/util/worker_pool.ts` 6. deck.gl off-thread attribute generation RFC: https://github.com/visgl/deck.gl/blob/master/dev-docs/RFCs/proposals/off-thread-attribute-generation-rfc.md, `ref-repos/visgl__deck.gl/dev-docs/RFCs/proposals/off-thread-attribute-generation-rfc.md` 7. ortho-earth persistent pool: https://github.com/kenjiyoshidahome2026-bit/ortho-earth/commit/103b127e, https://github.com/kenjiyoshidahome2026-bit/ortho-earth/commit/b036443d 8. Cesium WASM decoders in workers and splat rebuild debounce: https://github.com/CesiumGS/cesium/pull/13669 9. ortho-earth COPC reader workers reverted: https://github.com/kenjiyoshidahome2026-bit/ortho-earth/commit/93720c28, https://github.com/kenjiyoshidahome2026-bit/ortho-earth/commit/7ca1bba5 10. ortho-earth geopbf: https://github.com/kenjiyoshidahome2026-bit/ortho-earth/blob/main/packages/geopbf/README.md, `ref-repos/kenjiyoshidahome2026-bit__ortho-earth/packages/geopbf/README.md` 11. QGIS tile encoding pool: https://github.com/qgis/QGIS/commit/6439498f90 12. RustGIS createImageBitmap: https://github.com/opengeos/RustGIS/pull/630, https://github.com/opengeos/RustGIS/commit/f8f84ac 13. Cesium ImageBitmap revert: `ref-repos/CesiumGS__cesium/CHANGES.md`, https://github.com/CesiumGS/cesium/pull/7579, https://github.com/CesiumGS/cesium/pull/9627, https://github.com/CesiumGS/cesium/issues/12846 14. deck.gl loader-worker triangulation: https://github.com/visgl/deck.gl/blob/master/dev-docs/RFCs/proposals/loader-worker-triangulation.md, `ref-repos/visgl__deck.gl/dev-docs/RFCs/proposals/loader-worker-triangulation.md`, https://github.com/visgl/deck.gl/blob/master/docs/whats-new.md, https://github.com/visgl/deck.gl/commit/bacb7c24, https://github.com/visgl/deck.gl/commit/0275d745, https://github.com/visgl/deck.gl/commit/8746716d 15. MLT columnar decoding: https://github.com/maplibre/maplibre-tile-spec/blob/main/rust/mlt-wasm/README.md, `ref-repos/yuiseki__maplibre-tile-spec/rust/mlt-wasm/README.md`, `ref-repos/yuiseki__maplibre-tile-spec/docs/implementation-guide.md` 16. OpenLayers raster source in a worker: https://github.com/openlayers/openlayers/pull/3214 17. navara font worker: https://github.com/maplibre/navara/pull/464, https://github.com/maplibre/navara/commit/04a7e1f7, https://github.com/maplibre/navara/commit/d0b0d573 18. MapLibre GL JS setData coalescing: https://github.com/maplibre/maplibre-gl-js/commit/fa95c19c0, https://github.com/maplibre/maplibre-gl-js/commit/725f2666b, https://github.com/maplibre/maplibre-gl-js/commit/e5497caf1 19. MapLibre GL JS worker pipeline: `ref-repos/maplibre__maplibre-gl-js/developer-guides/life-of-a-tile.md`, `ref-repos/maplibre__maplibre-gl-js/ARCHITECTURE.md`, https://github.com/mapbox/mapbox-gl-js/pull/7124 20. MapLibre GL JS global placement: https://github.com/maplibre/maplibre-gl-js/commit/135561825, https://github.com/maplibre/maplibre-gl-js/commit/aefb3ee1f, https://github.com/maplibre/maplibre-gl-js/commit/a1afdd253 21. MapLibre Native global placement: https://github.com/maplibre/maplibre-native/commit/c0cb210dd, https://github.com/maplibre/maplibre-native/commit/4c6f0e813, https://github.com/maplibre/maplibre-native/commit/18d735ccc 22. Tangram JS collision split: https://github.com/tangrams/tangram/commit/d4d4606c, https://github.com/tangrams/tangram/commit/8ff7ca80, https://github.com/tangrams/tangram/commit/3f50d8c9 23. MapLibre Native synchronousUpdate: https://github.com/maplibre/maplibre-native/pull/3968 24. OpenLayers worker-built point buffers: https://github.com/openlayers/openlayers/commit/65be90709, https://github.com/openlayers/openlayers/commit/732b383b5, https://github.com/openlayers/openlayers/commit/5c1191dd5 25. ortho-earth same-frame overlays: https://github.com/kenjiyoshidahome2026-bit/ortho-earth/issues/13 26. MapLibre GL JS StructArray: https://github.com/maplibre/maplibre-gl-js/commit/11c8bb562, https://github.com/maplibre/maplibre-gl-js/commit/7b638de40 27. deck.gl MVT decoding in a worker: https://github.com/visgl/deck.gl/pull/3935, https://github.com/uber-web/loaders.gl/issues/685 28. MapLibre GL JS transfer vs copy: https://github.com/maplibre/maplibre-gl-js/commit/c465272a3, https://github.com/maplibre/maplibre-gl-js/commit/0310542cd, https://github.com/maplibre/maplibre-gl-js/commit/e2dce3ba4 29. ortho-earth IndexedDB row-group cache: https://github.com/kenjiyoshidahome2026-bit/ortho-earth/commit/c5492461 30. Tangram tile slice: https://github.com/tangrams/tangram/blob/master/src/tile/tile.js, `ref-repos/tangrams__tangram/src/tile/tile.js` 31. MapLibre GL JS updates without round trips: https://github.com/maplibre/maplibre-gl-js/pull/6668, https://github.com/maplibre/maplibre-gl-js/issues/4364, https://github.com/maplibre/maplibre-gl-js/commit/ad813fad1, https://github.com/maplibre/maplibre-gl-js/commit/a30582601, https://github.com/maplibre/maplibre-gl-js/commit/c842c3bd6 32. MapLibre GL JS style to workers first: https://github.com/maplibre/maplibre-gl-js/commit/7351d2fbf 33. Tangram stringified messages: https://github.com/tangrams/tangram/commit/3d874d22, https://github.com/tangrams/tangram/commit/1621da0c, https://github.com/tangrams/tangram/commit/8d49f81a 34. MapLibre GL JS MVT as transfer format: https://github.com/maplibre/maplibre-gl-js/pull/6803, https://github.com/maplibre/vt-pbf/pull/16 35. navara fresh WASM memory views: https://github.com/maplibre/navara/commit/268b8aab, https://github.com/maplibre/navara/commit/4af8b7b0, https://github.com/maplibre/navara/commit/6087f0c1 36. Safari Transferables: `ref-repos/maplibre__maplibre-gl-js/CHANGELOG.md`, https://github.com/mapbox/mapbox-gl-js/pull/9003, https://github.com/maplibre/maplibre-gl-js/pull/3225 37. maplibre-rs web threading: https://github.com/maplibre/maplibre-rs/blob/main/docs/src/supported-platforms.md, https://github.com/maplibre/maplibre-rs/issues/45, https://github.com/maplibre/maplibre-rs/commit/7dc7346, https://github.com/maplibre/maplibre-rs/commit/b7e4c27 38. ortho-earth SharedArrayBuffer fallback: https://github.com/kenjiyoshidahome2026-bit/ortho-earth/blob/main/packages/ortho-core/fallback-ladder.md, `ref-repos/kenjiyoshidahome2026-bit__ortho-earth/packages/ortho-core/fallback-ladder.md` 39. OpenLayers lazy worker: https://github.com/openlayers/openlayers/commit/2f7a891b9, https://github.com/openlayers/openlayers/commit/ce2d95ee8 40. Cesium Draco in a worker: https://github.com/CesiumGS/cesium/pull/6191 41. MapLibre GL JS pool lifetime and liveness: https://github.com/maplibre/maplibre-gl-js/pull/8491, https://github.com/maplibre/maplibre-gl-js/issues/8461 42. reearth-papers render permits: https://github.com/reearth/reearth-papers/commit/2007668, https://github.com/reearth/reearth-papers/commit/eabf229, `ref-repos/reearth__reearth-papers/CONTRIBUTING.md` 43. MapLibre GL JS v6 setWorkerUrl: https://github.com/maplibre/maplibre-gl-js/pull/6254 44. Cesium inlined workers: https://github.com/CesiumGS/cesium/commit/07b105ca2b448643db5c2914a37a1583cc94440e, https://github.com/CesiumGS/cesium/commit/18e18021ee46dbf89526921038c34b2e89d0e8f3 45. Tangram Blob worker build: https://github.com/tangrams/tangram/blob/master/rollup.config.mjs, `ref-repos/tangrams__tangram/rollup.config.mjs`, https://github.com/tangrams/tangram/blob/master/build/intro.js 46. Cesium shared worker chunks: https://github.com/CesiumGS/cesium/commit/aabd27a76, https://github.com/CesiumGS/cesium/pull/8224 47. ortho-earth single worker entry and nested workers: https://github.com/kenjiyoshidahome2026-bit/ortho-earth/issues/12 48. deck.gl worker from a CDN: https://github.com/visgl/deck.gl/commit/8d824a4b 49. `tools/build-site.ts` (this repository) 50. MapLibre GL JS cancel and abort: https://github.com/maplibre/maplibre-gl-js/commit/b5c90201d, https://github.com/maplibre/maplibre-gl-js/commit/1d3d433d9, https://github.com/maplibre/maplibre-gl-js/commit/368a604f1, https://github.com/maplibre/maplibre-gl-js/commit/95c703172 51. MapLibre Native geometry tile worker: `ref-repos/maplibre__maplibre-native/docs/mdbook/src/design/geometry-tile-worker.md` 52. tangram-es source generation: https://github.com/tangrams/tangram-es/pull/400 53. tangram-es non-blocking workers: https://github.com/tangrams/tangram-es/commit/ea75fe42, https://github.com/tangrams/tangram-es/commit/d5db23bf, https://github.com/tangrams/tangram-es/commit/d544e564 54. MapLibre GL JS batched glyph requests: https://github.com/maplibre/maplibre-gl-js/commit/726af278a 55. QGIS label cancellation: `ref-repos/qgis__QGIS/ChangeLog` (2019-11-29 [pal] more exit points), https://github.com/qgis/QGIS/issues/32489 56. GRASS banded parallel rasters: https://github.com/OSGeo/grass/commit/251d7f0d9, https://github.com/OSGeo/grass/commit/c80cbf1f6, https://github.com/OSGeo/grass/commit/935447215 57. GRASS shared MASK: https://github.com/OSGeo/grass/issues/2809, https://github.com/OSGeo/grass/pull/2829 58. Marble decoder mutex: https://github.com/KDE/marble/commit/28f9362ea, https://github.com/KDE/marble/commit/5598411f1 59. GRASS shared flag: https://github.com/OSGeo/grass/commit/bf7d8b6f5 60. GRASS shared RNG: https://github.com/OSGeo/grass/pull/6480, https://github.com/OSGeo/grass/commit/c83afef6d, https://github.com/OSGeo/grass/commit/becef3d05 61. GRASS atomic accumulation: https://github.com/OSGeo/grass/commit/1f83c2d27 62. Work splitting: https://github.com/OSGeo/grass/commit/9843891c6, https://github.com/OSGeo/grass/commit/2e636fb7a, https://github.com/KDE/marble/commit/166159644 63. GRASS r.param.scale: https://github.com/OSGeo/grass/pull/7440 64. GRASS thread defaults: https://github.com/OSGeo/grass/pull/5731, https://github.com/OSGeo/grass/commit/110ff3393, https://github.com/OSGeo/grass/commit/9a61f34e0, https://github.com/OSGeo/grass/commit/0b4a96034 65. Tangram tile-to-worker assignment: https://github.com/tangrams/tangram/commit/8a4eb885, https://github.com/tangrams/tangram/commit/311fd022, https://github.com/tangrams/tangram/commit/7a4c4792 66. GRASS worker threads off by default: https://github.com/OSGeo/grass/commit/0fb91788d, https://github.com/OSGeo/grass/commit/3d1354013, https://github.com/OSGeo/grass/commit/3500fb0a8 67. `harness/runners/browser.ts` (this repository)