geo-engine / geoengine / 42983422653

88%

main: 88%

Build Type

push

github

Committed by web-flow

Commit Message

fix(backend): make the new raster cache safe and fast under concurrent queries (#1420)

* fix(backend): never hold a raster cache lock across an await

The in-memory cache store held tokio locks across awaits. A future that holds
one and is not polled again, e.g. in a buffered stream nobody reads, kept the
cache locked for all other queries. With std locks the compiler rejects holding
a guard across an await, and record_hit, clear and the LRU update are
synchronous.

* test(backend): bench queries through the raster cache operator

* perf(backend): look up and compute raster cache tiles concurrently

The raster cache operator handled one tile after another, so every miss
waited for the source query of the previous one. It now reads ahead 16 tiles,
like the GDAL sources, and keeps the tile order.

raster_cache bench, NDVI, median of 20 runs:
tiles_256 cold 14.3 -> 5.1 ms, warm 13.1 -> 6.4 ms
tiles_512 cold 25.4 -> 8.7 ms, warm 25.3 -> 9.3 ms

* perf(backend): store a raster cache entry once when inserts race

Callers that got the same result, e.g. from the GDAL pool's shared reads,
insert it at the same time. All of them passed the check for an existing
entry and compressed the value, and all but one threw it away. An insert now
claims its key before it compresses; a concurrent insert of the same key
returns.

* perf(backend): decompress cached tiles directly into their buffer

Decompressing into bytes and copying them to the pixel type allocated every
buffer twice and read pixels through a possibly misaligned slice.

* perf(backend): load raster cache entries without copying them

A hit copied the whole compressed tile on the task that polls the query, only
to move it into the thread that decompresses it. That thread now shares the
entry, which also keeps it from being evicted meanwhile.

* feat(backend): count hits, misses and joined inserts of the raster cache

(cherry picked from commit 21418f1a2)

* refactor(backend)... (continued)

Coverage Stats

157 of 182 new or added lines in 3 files covered. (86.26%)

5 existing lines in 2 files now uncovered.

137887 of 156687 relevant lines covered (88.0%)

536246.09 hits per line