Repository navigation
Fetch COG tiles concurrently in HTTP path to mask RTT - #1487
Merged
Merged
Conversation
_read_cog_http walked tiles one at a time, blocking on each range request before sending the next. Over a 50 ms RTT link, a 100-tile COG paid 5 s in round trips before any data flowed. Add _HTTPSource.read_ranges, which takes a list of (offset, length) pairs and fetches them through a ThreadPoolExecutor (default 8 workers, override via XRSPATIAL_COG_HTTP_WORKERS). _read_cog_http collects every tile range up front, fetches them in parallel, then decodes and places each tile in input order. read_range is unchanged so other call sites are unaffected. Local file reads do not go through this path. Tests: order preservation, empty-list handling, single-request fast path, concurrency speedup against an artificial-latency fake source, and a round-trip correctness test against a local http.server.
The previous test_read_ranges_concurrency_masks_latency required 20 fake-latency requests at 50ms each to finish under 500ms. On a busy macOS CI runner the run took 519ms — 4% over the budget — and tripped fail-fast across the matrix. Track max-in-flight count in _FakeHTTPSource and assert >= 2. That proves the threadpool dispatches concurrently without depending on wall-clock timing.
This was referenced May 15, 2026
brendancol
added a commit
to brendancol/xarray-spatial
that referenced
this pull request
May 16, 2026
The HTTP COG read path fetches tiles concurrently via ``read_ranges_coalesced`` (xarray-contrib#1480, xarray-contrib#1487) and then decoded them sequentially in a Python for-loop. Local ``_read_tiles`` already parallelises decode via ``ThreadPoolExecutor`` whenever ``tile_pixels >= 64 * 1024`` and ``n_tiles > 1`` (see _reader.py:2017); the HTTP path was structurally similar but never adopted the same gate, so wide windowed COG reads with many tiles left deflate/zstd decoding single-threaded after a parallel fetch. Mirror the local-file path's threshold and pool here. Codec extensions release the GIL inside their C implementations, so a ``ThreadPoolExecutor`` overlaps decode work across cores. The placement loop that writes pixels into ``result`` stays serial to avoid contended writes. Pass 10 of the geotiff performance sweep. Tests in test_cog_http_parallel_decode_2026_05_15.py: - parallel-branch end-to-end correctness (deflate, tile_size=256) - serial-branch end-to-end correctness (tile_size=128, single tile) - ThreadPoolExecutor instantiation above the threshold - single-tile path skips the decode pool - _decode_strip_or_tile invoked exactly once per placement
brendancol
added a commit
that referenced
this pull request
May 16, 2026
…) (#1981) * geotiff: parallelise tile decode in _fetch_decode_cog_http_tiles The HTTP COG read path fetches tiles concurrently via ``read_ranges_coalesced`` (#1480, #1487) and then decoded them sequentially in a Python for-loop. Local ``_read_tiles`` already parallelises decode via ``ThreadPoolExecutor`` whenever ``tile_pixels >= 64 * 1024`` and ``n_tiles > 1`` (see _reader.py:2017); the HTTP path was structurally similar but never adopted the same gate, so wide windowed COG reads with many tiles left deflate/zstd decoding single-threaded after a parallel fetch. Mirror the local-file path's threshold and pool here. Codec extensions release the GIL inside their C implementations, so a ``ThreadPoolExecutor`` overlaps decode work across cores. The placement loop that writes pixels into ``result`` stays serial to avoid contended writes. Pass 10 of the geotiff performance sweep. Tests in test_cog_http_parallel_decode_2026_05_15.py: - parallel-branch end-to-end correctness (deflate, tile_size=256) - serial-branch end-to-end correctness (tile_size=128, single tile) - ThreadPoolExecutor instantiation above the threshold - single-tile path skips the decode pool - _decode_strip_or_tile invoked exactly once per placement * geotiff: DRY parallel-decode threshold + helper across local/HTTP tile paths Review nits on the parallel-HTTP-decode change: - Extract the 64K-pixel cutoff into a module-level _PARALLEL_DECODE_PIXEL_THRESHOLD constant. The same magic number lived in _read_tiles and _fetch_decode_cog_http_tiles; the only docstring describing why it was chosen sat next to the local path. - Lift _decode_one out of the parallel branch in _fetch_decode_cog_http_tiles so the serial fallback calls the same helper instead of inlining a duplicate _decode_strip_or_tile call. Mirrors how _read_tiles defines _decode_one once and reuses it for both paths. - Drop the ThreadPoolExecutor-as-_Pool alias for parity with the local path, which imports ThreadPoolExecutor by its real name. No behaviour change. Existing 268 COG/HTTP and 734 tile/reader tests still pass; the 6 pre-existing failures (predictor2 big-endian GPU, tile_size=4 validator) predate this PR per the prior sweep notes.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #1480.
Summary
_read_cog_httpwalked tiles one at a time, blocking on each range request before sending the next. Over a 50 ms RTT link, a 100-tile COG paid 5 s in round trips before any data flowed.This PR adds
_HTTPSource.read_ranges, which takes a list of(offset, length)pairs and fetches them through aThreadPoolExecutor. Pool size defaults to 8 and can be overridden withXRSPATIAL_COG_HTTP_WORKERS._read_cog_httpnow collects every tile range up front, fetches them in parallel, then decodes and places each tile in input order._HTTPSource.read_rangeis unchanged, so no other call sites are affected. Local file reads do not go through this path.Measured on the test fake source (20 requests, 50 ms each, 8 workers): sequential bound is 1.00 s, concurrent fetch took 0.15 s, a 6.6x speedup.
Test plan
pytest xrspatial/geotiff/tests/test_cog_http_concurrent.py-- 6 new tests cover ordering, empty list, single-range fast path, concurrency speedup vs an artificial-latency fake, and round-trip correctness against a localhttp.server.pytest xrspatial/geotiff/tests/test_cog.py xrspatial/geotiff/tests/test_reader.py-- existing 32 tests still pass.