Repository navigation
Description
Summary
A graph_batch() has a fixed cost that does not depend on how many edges it gets. With the default arguments, a batch of one edge takes about 2 ms and allocates about 41 MiB on the JVM wheel, and about 2.5 ms on the native wheel. With batch_size=100 and chunk_cache_capacity=1000 the same call takes about 0.2 ms and allocates 0.6 MiB. A workload that opens a batch per request or per small group of edges (an agent writing a few relationships per turn, a stream writer) pays this on every batch, and the garbage makes the JVM heap grow (see the NMT note below). bulk-loads.adoc and the graph_batch() docstring describe both arguments but not this cost, so nothing tells a caller of small batches to change them.
Reproduction
Timing, both wheels:
import os
import tempfile
import time
import arcadedb
print(arcadedb.__backend__, arcadedb.__version__)
db = arcadedb.create_database(os.path.join(tempfile.mkdtemp(), "db"))
db.command("sql", "CREATE VERTEX TYPE V")
db.command("sql", "CREATE EDGE TYPE E")
with db.transaction():
for _ in range(1000):
db.command("sql", "CREATE VERTEX V")
rids = [str(r.get_rid()) for r in db.query("sql", "SELECT FROM V")]
def run(label, **options):
def one():
with db.graph_batch(**options) as batch:
batch.new_edge(rids[0], "E", rids[1])
for _ in range(200): # warm up
one()
start = time.perf_counter()
for _ in range(1000):
one()
micros = (time.perf_counter() - start) / 1000 * 1e6
print(f"{label:45} {micros:7.0f} us per one-edge graph_batch")
run("default")
run("batch_size=100, chunk_cache_capacity=1000", batch_size=100, chunk_cache_capacity=1000)native 26.11.1.post4
default 2474 us per one-edge graph_batch
batch_size=100, chunk_cache_capacity=1000 220 us per one-edge graph_batch
jvm 26.12.1.dev0
default 2235 us per one-edge graph_batch
batch_size=100, chunk_cache_capacity=1000 250 us per one-edge graph_batch
Allocation, JVM wheel (bytes the JVM allocates per graph_batch, summed over its threads with ThreadMXBean.getThreadAllocatedBytes):
import os
import tempfile
import jpype
import arcadedb
db = arcadedb.create_database(os.path.join(tempfile.mkdtemp(), "db"))
db.command("sql", "CREATE VERTEX TYPE V")
db.command("sql", "CREATE EDGE TYPE E")
with db.transaction():
for _ in range(1000):
db.command("sql", "CREATE VERTEX V")
rids = [str(r.get_rid()) for r in db.query("sql", "SELECT FROM V")]
bean = jpype.JClass("java.lang.management.ManagementFactory").getThreadMXBean()
def allocated():
return sum(max(0, int(bean.getThreadAllocatedBytes(t))) for t in bean.getAllThreadIds())
def run(label, edges, **options):
def one():
with db.graph_batch(**options) as batch:
if edges:
batch.new_edge(rids[0], "E", rids[1])
for _ in range(200):
one()
before = allocated()
for _ in range(1000):
one()
print(f"{label:45} {(allocated() - before) / 1000 / 1024:9.1f} KiB per graph_batch")
run("no edge, default", False)
run("one edge, default", True)
run("one edge, chunk_cache_capacity=1000", True, chunk_cache_capacity=1000)
run("one edge, batch_size=100", True, batch_size=100)
run("one edge, both", True, batch_size=100, chunk_cache_capacity=1000)no edge, default 8248.6 KiB per graph_batch
one edge, default 41568.1 KiB per graph_batch
one edge, chunk_cache_capacity=1000 8829.9 KiB per graph_batch
one edge, batch_size=100 33374.2 KiB per graph_batch
one edge, both 636.5 KiB per graph_batch
macOS arm64 (16 cores), Python 3.12, JDK 25 runtime of the JVM wheel; the JVM wheel was built from main at 0c3c7ec. Timings are single runs of 1,000 batches after 200 warm-up batches; the ratio held in four repeated runs (default 1.4 to 2.2 ms, tuned 0.19 to 0.30 ms).
Notes
- An async-profiler allocation profile of the default one-edge loop puts most of the allocated bytes under
com.arcadedb.graph.GraphBatch.<init>(long[], int[] and HashMap node arrays), which is engine code. The bindings passbatch_sizeandchunk_cache_capacitythrough unchanged, so the default sizes come from the engine. - Heap effect, JVM wheel: with
-XX:NativeMemoryTracking=summary, 9,000 one-edge batches in a loop grew the committed Java heap by 7.6 GiB (the heap does not leak: used heap after a full GC stayed flat over 20,000 batches in a separate run), because the allocation rate makes G1 expand the heap. - Whether the engine should size these structures lazily is a question for ArcadeData/arcadedb. What belongs here is to say it: a short "many small batches" paragraph in
bulk-loads.adocand in thebatch_sizeandchunk_cache_capacityentries of thegraph_batch()docstring. I can send that as a PR.
Activity
shaunpatterson commentedon Nov 21, 2026
Two more rows of the sweep, JVM wheel, same script:
expected_edge_count=1does not change the cost (41,565 KiB and about 1.6 ms per one-edge batch, the same as the default), andchunk_cache_capacity=1000alone takes the allocation from 41,569 KiB to 8,832 KiB with the time unchanged (about 1.4 ms), so the chunk caches account for about 32 MiB of the 41 MiB;batch_size=100alone leaves 33 MiB. Both arguments are needed to reach 0.6 MiB and 0.2 ms.