You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
JVM wheel: Arrow vertex loads with a vector column use 1.3 to 1.4 times the memory of the 26.11.1 route (speed fixed in post2) #348
While fixing the JVM wheel's Arrow vertex loads for 26.11.1.post2 (tracking in #347) I measured memory next to speed, and the memory did not improve with the speed.
Shape: insert_arrow of a vertex type with one 96-wide float32 vector column, 10 calls of 10,000 rows, JVM wheel, released 26.11.1 (A) against the post2 candidate (B), write-ahead log forced the same way on both arms, 20 clean pairs per cell, interval = 95 percent bootstrap of the median of per-round ratios (B/A above 1 is more memory for B).
WAL
peak RSS B/A
RssAnon B/A
JVM heap used at the end, A / B (MiB)
on
1.273 [1.237, 1.341]
1.314 [1.282, 1.395]
164.2 / 308.0
off
1.387 [1.326, 1.460]
1.448 [1.377, 1.535]
161.7 / 299.9
Speed for the same cells, so the trade is visible: the WAL-off cell is a WIN at B/A 0.154 [0.150, 0.163] (time ratio, B takes that fraction of A). The other cells point the same way but were not decided by the pre-registered rule (a rerun under a written amendment is queued). This issue is about memory only: the speed change did not come with a memory change in the same direction.
Where it goes: the post2 change removed the per-row Python conversion of the vector column (one float32 matrix per batch crosses to Java instead of one boxed list per row), which an untimed profile put at about three quarters of the call. The rest of the excess is probably on the Java side (the per-row float[] copy in FloatVectors.Matrix.row, the document build, the transaction), but I have not profiled that yet.
I will look at the Java side next. If someone who knows the columnar insert path sees where the 96-float rows are held, I would take that pointer.
The rerun of the A/B of the vector-conversion fix has finished (the released 26.11.1 JVM wheel against the fix, which 26.11.1.post2 carries). I wrote an amendment to the study rule after the first pass and disclosed it in the study README; this rerun is scored under it. Arrow vertex loads with a float32 vector column, the write-ahead log on and off the same way on both arms, 12 clean pairs per figure, B/A of the medians with the 95 percent interval (below 1.00 means the fix is faster):
figure
time B/A
peak memory B/A
j_e2_ar_off
0.234 [0.213, 0.254]
1.026 [0.964, 1.115]
j_e2_ar_on
0.258 [0.230, 0.333]
1.125 [1.004, 1.197]
j_l3d_ar_off
0.157 [0.149, 0.174]
1.420 [1.353, 1.485]
j_l3d_ar_on
0.180 [0.175, 0.188]
1.323 [1.277, 1.447]
j_sc_ar_off
0.341 [0.296, 0.356]
1.200 [1.131, 1.345]
j_sc_ar_on
0.334 [0.302, 0.359]
1.005 [0.912, 1.223]
The fix is faster in all six figures. Memory is not improved: on the vector shape (l3d) the fix stays 1.3 to 1.4 times above the released route, as in the first pass, and the other two shapes are mixed. This issue stays open for the memory side. I have not looked at the cause yet. If you want the memory to be the next fix, tell me here and I will profile the conversion.
Activity
tae898 commentedon Nov 10, 2026
The rerun of the A/B of the vector-conversion fix has finished (the released
26.11.1JVM wheel against the fix, which26.11.1.post2carries). I wrote an amendment to the study rule after the first pass and disclosed it in the study README; this rerun is scored under it. Arrow vertex loads with a float32 vector column, the write-ahead log on and off the same way on both arms, 12 clean pairs per figure, B/A of the medians with the 95 percent interval (below 1.00 means the fix is faster):j_e2_ar_offj_e2_ar_onj_l3d_ar_offj_l3d_ar_onj_sc_ar_offj_sc_ar_onThe fix is faster in all six figures. Memory is not improved: on the vector shape (
l3d) the fix stays 1.3 to 1.4 times above the released route, as in the first pass, and the other two shapes are mixed. This issue stays open for the memory side. I have not looked at the cause yet. If you want the memory to be the next fix, tell me here and I will profile the conversion.Docs: say that the JVM Arrow vector load is faster and uses more memo…