ci(benchmark): aggressive shard rebalancing for 5-minute target

Split benchmarks into 15 shards to ensure each completes within
5 minutes. Separated slow vs non-slow tests for expensive operations
(all, items, keys) using the slow marker.

Shard distribution (estimated times):
- get: 25 tests (~3m)
- count: 25 tests (~3m)
- has: 31 tests (~3m)
- find-one-delete: 24 tests (~2m)
- write-compact: 32 tests (~5m)
- load: 19 tests (~4m)
- reload: 19 tests (~4m)
- keys-ci: 12 tests (~2.5m)
- keys-slow: 13 tests (~2.5m)
- all-ci: 12 tests (~3m)
- all-slow: 13 tests (~3.5m)
- items-ci: 12 tests (~3m)
- items-slow: 13 tests (~3.5m)
- find-high: 19 tests (~3m)
- find-other: 24 tests (~4m)

Memory tests excluded for now due to memray profiling overhead.

Author: tbhb

.github/workflows/benchmark.yml

@@ -31,47 +31,55 @@ jobs:
   benchmark:
     name: "Benchmark (${{ matrix.shard }})"
     runs-on: ubuntu-latest
-    timeout-minutes: 6
+    timeout-minutes: 5
     permissions:
       contents: read
       id-token: write
     strategy:
       fail-fast: false
       matrix:
         include:
-          # ~38 tests each for load/reload
-          - shard: "load-reload"
-            filter: "TestBenchLoad or TestBenchReload"
-          # 25 tests
+          # Fast tests (~7s/test) - can have more tests per shard
           - shard: "get"
             filter: "TestBenchGet"
-          # 25 tests
-          - shard: "all"
-            filter: "TestBenchAll"
-          # 43 tests (find_high=19, find_other=24)
-          - shard: "find"
-            filter: "TestBenchFind"
-          # 18 tests
-          - shard: "find-one"
-            filter: "TestBenchFindOne"
-          # 32 tests
-          - shard: "write-compact"
-            filter: "TestBenchPut or TestBenchBatchWrite or TestBenchCompact"
-          # 31 tests
-          - shard: "keys-delete"
-            filter: "TestBenchKeys or TestBenchDelete"
-          # 25 tests
-          - shard: "items"
-            filter: "TestBenchItems"
-          # 25 tests
           - shard: "count"
             filter: "TestBenchCount"
-          # 31 tests
           - shard: "has"
             filter: "TestBenchHas"
-          # 18 tests
-          - shard: "memory"
-            filter: "TestMemory"
+          - shard: "find-one-delete"
+            filter: "TestBenchFindOne or TestBenchDelete"
+          # Moderate tests (~9s/test)
+          - shard: "write-compact"
+            filter: "TestBenchPut or TestBenchBatchWrite or TestBenchCompact"
+          # I/O-heavy tests (~12s/test)
+          - shard: "load"
+            filter: "TestBenchLoad"
+          - shard: "reload"
+            filter: "TestBenchReload"
+          - shard: "keys-ci"
+            filter: "TestBenchKeys"
+            marker: "benchmark and not slow"
+          - shard: "keys-slow"
+            filter: "TestBenchKeys"
+            marker: "benchmark and slow"
+          # Expensive tests (~15s/test) - split by slow marker
+          - shard: "all-ci"
+            filter: "TestBenchAll"
+            marker: "benchmark and not slow"
+          - shard: "all-slow"
+            filter: "TestBenchAll"
+            marker: "benchmark and slow"
+          - shard: "items-ci"
+            filter: "TestBenchItems"
+            marker: "benchmark and not slow"
+          - shard: "items-slow"
+            filter: "TestBenchItems"
+            marker: "benchmark and slow"
+          # Find tests (~10s/test)
+          - shard: "find-high"
+            filter: "TestBenchFind and not TestBenchFindOne and test_find_high_selectivity"
+          - shard: "find-other"
+            filter: "TestBenchFind and not TestBenchFindOne and not test_find_high_selectivity"
 
     steps:
       - name: Checkout code
@@ -87,4 +95,4 @@ jobs:
         uses: CodSpeedHQ/action@346a2d8a8d9d38909abd0bc3d23f773110f076ad # v4.4.1
         with:
           mode: simulation
-          run: uv run pytest -m benchmark -k "${{ matrix.filter }}" --codspeed
+          run: uv run pytest -m "${{ matrix.marker || 'benchmark' }}" -k "${{ matrix.filter }}" --codspeed