Durability of the databases measured
For these performance figures, every database is measured in persistent mode (data written to disk, replayed on restart).
Databases measured
| Database | Version measured | Edition or licence | Protocol | Measurements |
|---|---|---|---|---|
| gDown | main @ ac35131, 2026-10-23 | JVM build, benchmark settings | Bolt | workloads |
| Neo4j | 5.26.31 community | Community (GPLv3) | Bolt | workloads |
| Memgraph | 3.13.1 | Community (BSL 1.1) | Bolt | workloads |
| FalkorDB | 6.0.2 on Redis 8.10.2 | SSPL 1.0 | RESP | workloads |
| ArcadeDB | 26.10.1 | Apache 2.0 | Bolt | workloads |
| TuGraph | 4.5.1 | Community (Apache 2.0) | Bolt | workloads |
| Apache AGE | 1.8.0 on PostgreSQL 18.6 (Debian 18.6-1.pgdg13+2) | Apache 2.0 (PostgreSQL extension) | JDBC | workloads |
How the benchmark runs
| Property | Value |
|---|---|
| Architecture | arm64 |
| Machine | Apple M4 Max, 16 logical CPUs, 64 GB RAM; Docker 29.9.2 |
| Container limits | 4 CPUs, 8 GB of memory per container |
| Graph of the workloads | 100,000 nodes, 1,000,000 relationships, generated with a fixed seed |
| Rounds | 3 full rounds; every figure is the median of the rounds |
| Clients | One Java client on the host; 1, 4, 8 concurrent clients depending on the workload |
| Protocol | The transport of each database (table above): Bolt on both sides for gDown and Neo4j; a warm-up round before each workload, not counted |
| Dates | 2026-10-24 – 2026-10-24 |
Startup, memory, CPU and size
Without index
| Workload | gDown | Neo4j | Memgraph | FalkorDB | ArcadeDB | TuGraph | Apache AGE |
|---|---|---|---|---|---|---|---|
| Startup to the first answer | 0.8 | 4.0 | 0.6 | 0.2 | 1.5 | 0.5 | 0.9 |
| CPU average (100 % = one core) | 100 | 180 | 234 | 100 | 102 | 104 | 109 |
| CPU peak | 274 | 385 | 408 | 161 | 207 | 404 | 244 |
| Peak memory | 899 | 4,805 | 542 | 426 | 6,415 | 381 | 476 |
| Data on disk | 176 | 836 | 389 | 39 | 120 | 173 | 529 |
With indexes on the identifiers
| Workload | gDown | Neo4j | Memgraph | FalkorDB | ArcadeDB | TuGraph | Apache AGE |
|---|---|---|---|---|---|---|---|
| Startup to the first answer | 0.8 | 4.0 | 0.6 | 0.2 | 1.5 | 0.5 | 0.9 |
| CPU average (100 % = one core) | 105 | 179 | 224 | 99 | 102 | 104 | 109 |
| CPU peak | 279 | 385 | 410 | 160 | 207 | 404 | 244 |
| Peak memory | 823 | 4,807 | 506 | 429 | 6,415 | 381 | 476 |
| Data on disk | 181 | 837 | 389 | 39 | 120 | 173 | 529 |
Startup: from docker run to the first answer to RETURN 1, median of several attempts, measured once and shown in both configurations; the other rows cover the whole run of the workloads. CPU and disk are not ranked: a database that writes nothing durable has 0 MB on disk.
Complex reads: median latency — LDBC SNB Interactive
The 29 queries of the LDBC Social Network Benchmark Interactive workload (14 complex reads, 7 short reads, 8 updates; Apache 2.0, unchanged) all run on a generated graph with the SNB schema.
- 29 / 29 queries run: 14 complex reads, 7 short reads, 8 updates
- 3 / 3 members of a 3-node cluster return the same row counts for the 21 reads
- 18 ms median write while the leader takes 1,365 acknowledged writes (3,000 persons)
- 0 ms median read on a follower during those writes (95th percentile: 30 ms)
All databases, 3,000 persons
| Workload | gDown | Neo4j | Memgraph | FalkorDB | ArcadeDB | TuGraph | Apache AGE |
|---|---|---|---|---|---|---|---|
| complex-1 | 2.9 | 7.4 | – | – | 10.0 | – | – |
| complex-2 | 1.1 | 2.4 | 1.6 | 211.9 | 2.3 | – | – |
| complex-3 | 11.3 | 9.8 | 10.3 | – | 17.7 | – | – |
| complex-4 | 0.9 | 1.9 | 0.7 | 0.6 | 1.9 | – | 131.5 |
| complex-5 | 7.5 | 9.5 | 4.0 | 2.2 | 15966.5 | – | – |
| complex-6 | 3.3 | 5.3 | 2.1 | – | 8.7 | – | – |
| complex-7 | 0.7 | 2.0 | 0.7 | – | – | – | – |
| complex-8 | 0.6 | 1.5 | 0.7 | 0.5 | 1.3 | – | – |
| complex-9 | 5.2 | 6.0 | 16.2 | – | 16.5 | – | – |
| complex-10 | 1.0 | 4.9 | – | – | – | – | – |
| complex-11 | 0.9 | 2.2 | 0.9 | 0.5 | 2.8 | – | – |
| complex-12 | 1.4 | 2.5 | 4.0 | – | 2.7 | – | – |
| complex-13 | 0.5 | 1.6 | – | – | 0.7 | – | – |
| complex-14 | 1.0 | 2.0 | – | – | 2002.9 | – | – |
A dash: the database refuses the query, exceeds its timeout, or returns a different number of rows than Neo4j (its time would mean nothing). Case by case:
All databases, 30,000 persons
- complex-1 : gDown 12.9 ms, Neo4j 51.5 ms
- complex-2 : gDown 2.8 ms, Neo4j 2 ms
- complex-3 : gDown 15.1 ms, Neo4j 13.3 ms
- complex-4 : gDown 1.6 ms, Neo4j 1.7 ms
- complex-5 : gDown 19.5 ms, Neo4j 16.9 ms
- complex-6 : gDown 6.7 ms, Neo4j 7 ms
- complex-7 : gDown 1.4 ms, Neo4j 1.4 ms
- complex-8 : gDown 0.8 ms, Neo4j 1 ms
- complex-9 : gDown 15.1 ms, Neo4j 8.5 ms
- complex-10 : gDown 4.4 ms, Neo4j 4.2 ms
- complex-11 : gDown 1.9 ms, Neo4j 1.8 ms
- complex-12 : gDown 2 ms, Neo4j 2.4 ms
- complex-13 : gDown 0.9 ms, Neo4j 1 ms
- complex-14 : gDown 4 ms, Neo4j 3.3 ms
Median over 5 persons after a warm-up round, one run per database, in containers limited as in the set-up table; every database is reached over its own transport (Bolt on both sides for gDown and Neo4j: the round trip is counted on both sides). “< 1 ms” means that the median rounds to zero milliseconds. Short reads: median under 1 ms for gDown (0.5 to 0.7 ms) against 0.7 to 1 ms for Neo4j, same protocol on both sides.
At 3,000 persons (90,000 nodes and 470,000 relationships, loaded through Raft in 16 s), medians stay under 60 ms. Two optimisations found by this test: complex-1 went from 0.6 s to 34 ms, complex-14 from 2 s to 8 ms.
Not official LDBC results: these figures are derived from the LDBC SNB Interactive workload and have not been audited by the LDBC (Linked Data Benchmark Council). The data is generated by the test, not by the LDBC Datagen: the times are not comparable with published LDBC results. One run per configuration. The test also found two defects, now fixed.
Source: gluonify-gdown, README and the reports of LdbcSnbTest and LdbcClusterTest.
Cypher coverage (the openCypher TCK on every base) is on the comparison page.
Latency and throughput, workload by workload
Same graph, same queries and the same machine for every database of the table above: the figures below are the median over the rounds of the set-up table, measured on arm64.
Latency is the median (p50) in milliseconds, lower is better; the workloads with several clients are in requests per second, higher is better. The best value of each row is in bold. Nothing is tuned beyond the defaults of each image; gDown runs on the JVM with a 6 GB heap and a 2 GB page cache, Neo4j with a 4 GB heap and a 2 GB page cache. Each run starts from a fresh container, loads the graph in UNWIND batches, checks that the databases agree on the results, then measures.
Without index
| Workload | gDown | Neo4j | Memgraph | FalkorDB | ArcadeDB | TuGraph | Apache AGE |
|---|---|---|---|---|---|---|---|
| Read by id | 0.30 | 0.72 | 0.29 | 0.16 | 0.35 | 0.46 | 0.23 |
| Read by property | 3.36 | 8.22 | 6.90 | 2.44 | 0.33 | 0.44 | 0.24 |
| Range on a property (100 values) | 3.14 | 10.7 | 11.6 | 3.20 | 24.1 | 75.8 | 0.24 |
| Count a label | 0.35 | 0.61 | 3.86 | 0.18 | 0.41 | 98.2 | 8.59 |
| 1 hop | 0.28 | 0.51 | 0.28 | 0.17 | 0.33 | 0.50 | 0.27 |
| 2 distinct hops | 0.33 | 0.56 | 0.30 | 0.20 | 0.40 | 0.60 | 0.72 |
| 3 distinct hops | 0.84 | 1.14 | 0.47 | 0.34 | 1.05 | 1.38 | 5.63 |
| Group by + top 10 | 3.27 | 9.88 | 31.1 | 15.7 | 8.65 | 145 | 61.4 |
| Filter + aggregate | 4.06 | 11.3 | 9.72 | 3.56 | 5.52 | 77.7 | 24.2 |
| Count a pattern (all KNOWS between persons) | 15.7 | 212 | 59.1 | 31.4 | 100 | 837 | 46.2 |
| Hot reads + a full traversal per 100 | 0.26 | 0.55 | 0.31 | 0.17 | 0.39 | 0.50 | 0.29 |
| First 10 rows of a large result | 0.28 | 0.58 | 0.37 | 0.20 | 0.57 | 0.53 | 0.19 |
| Write: 1 node per request | 0.40 | 0.55 | 0.42 | 0.34 | 0.51 | 0.45 | 0.25 |
| Writes while a long scan runs | 0.49 | 0.54 | 0.48 | 35.1 | 0.60 | 0.48 | 0.27 |
| 8 clients: 1 hop | 16,313 | 8,438 | 8,863 | 29,156 | 10,884 | 4,723 | 16,715 |
| 4 clients: writes | 6,568 | 5,242 | 3,524 | 4,851 | 3,452 | 3,618 | 9,302 |
ArcadeDB (indexed) is always measured with an index: its internal identifiers rule out the lookup by id() used for the other columns. Its column is here so that nothing is missing, but its ratios without index compare an index with a scan: see the table with indexes for the like-for-like comparison.
With indexes on the identifiers
| Workload | gDown | Neo4j | Memgraph | FalkorDB | ArcadeDB | TuGraph | Apache AGE |
|---|---|---|---|---|---|---|---|
| Read by id | 0.29 | 0.72 | 0.29 | 0.16 | 0.35 | 0.46 | 0.23 |
| Read by property | 0.27 | 0.50 | 0.27 | 0.18 | 0.33 | 0.44 | 0.24 |
| Range on a property (100 values) | 0.30 | 0.48 | 0.29 | 0.18 | 24.1 | 75.8 | 0.24 |
| Count a label | 0.26 | 0.47 | 3.64 | 0.15 | 0.41 | 98.2 | 8.59 |
| 1 hop | 0.26 | 0.46 | 0.27 | 0.16 | 0.33 | 0.50 | 0.27 |
| 2 distinct hops | 0.31 | 0.51 | 0.29 | 0.18 | 0.40 | 0.60 | 0.72 |
| 3 distinct hops | 0.76 | 0.99 | 0.44 | 0.31 | 1.05 | 1.38 | 5.63 |
| Group by + top 10 | 3.28 | 9.52 | 31.9 | 15.0 | 8.65 | 145 | 61.4 |
| Filter + aggregate | 4.05 | 15.4 | 9.73 | 3.53 | 5.52 | 77.7 | 24.2 |
| Count a pattern (all KNOWS between persons) | 15.7 | 209 | 61.6 | 31.4 | 100 | 837 | 46.2 |
| Hot reads + a full traversal per 100 | 0.26 | 0.53 | 0.31 | 0.17 | 0.39 | 0.50 | 0.29 |
| First 10 rows of a large result | 0.27 | 0.56 | 0.32 | 0.20 | 0.57 | 0.53 | 0.19 |
| Write: 1 node per request | 0.39 | 0.57 | 0.43 | 0.32 | 0.51 | 0.45 | 0.25 |
| Writes while a long scan runs | 0.50 | 0.54 | 0.49 | 35.0 | 0.60 | 0.48 | 0.27 |
| 8 clients: 1 hop | 16,365 | 8,427 | 8,940 | 28,988 | 10,884 | 4,723 | 16,715 |
| 4 clients: writes | 6,530 | 5,089 | 3,544 | 5,112 | 3,452 | 3,618 | 9,302 |
Loading the graph
Without index
| Workload | gDown | Neo4j | Memgraph | FalkorDB | ArcadeDB | TuGraph | Apache AGE |
|---|---|---|---|---|---|---|---|
| Load time | 7.9 | 8.7 | 3.3 | 5.1 | 14.4 | 102.3 | 98.9 |
| Loading rate | 138,434 | 126,238 | 328,444 | 214,623 | 76,137 | 10,749 | 11,118 |
With indexes on the identifiers
| Workload | gDown | Neo4j | Memgraph | FalkorDB | ArcadeDB | TuGraph | Apache AGE |
|---|---|---|---|---|---|---|---|
| Load time | 10.2 | 10.9 | 4.6 | 40.4 | 14.4 | 102.3 | 98.9 |
| Loading rate | 108,100 | 100,503 | 239,107 | 27,242 | 76,137 | 10,749 | 11,118 |
What to take from it
- FalkorDB is the fastest on short operations (reads by id, hops, single-node writes) and on throughput with several clients, with the shortest client path (RESP). Look at “writes while a long scan runs”: its writes wait behind a long read; it writes nothing durable by default and its loading with indexes is the slowest.
- gDown is faster than Neo4j on most workloads (geometric mean below 1 in both configurations), loads faster and keeps its data in far less disk space. It is level or slower on the first rows of a large result and on single-node writes: see First 10 rows and Write: 1 node.
- Against Memgraph, gDown wins on grouping, label counts and throughput with 8 clients, and loses on short reads, hops and writes. Memgraph does not force its commits to disk each time, unlike gDown and Neo4j.
- ArcadeDB, always measured with an index, is competitive on reads but slow on ranges and on writes with several clients.
- Memory: the databases that keep everything in RAM (Memgraph, FalkorDB) use the least. gDown and Neo4j run here with generous heaps, and gDown needs far less with a smaller heap for the same latencies (see the native measure above).
What this comparison does not say
- Durability differs: gDown and Neo4j sync to disk at each commit; Memgraph only every 100,000 transactions; FalkorDB writes nothing durable by default; the settings of the other databases were not checked. The write lines therefore do not compare the same guarantee.
- The databases that speak Bolt share the same Java driver; FalkorDB is measured with a minimal RESP client written for the benchmark, which shortens its path on reads (see the transport column).
- ArcadeDB is always measured with an index, in both tables (its internal identifiers rule out the lookup by id()); a database without a value in a configuration was not measured in it.
- The graph fits in memory for all of them. A graph larger than RAM, replication, failover, vector search, graph algorithms and reads during concurrent writes are not measured. One node of each database.
- ArcadeDB runs with an unbounded heap: its memory figure is not comparable. Its image is a development snapshot.
- A machine that was not idle, a few rounds: count on ±10 % on every cell.
Source: gluonify-gdown, the raw results of each round (docs/bench) and scripts/bench-compare.py, assembled by a script of this site; the benchmark script is in the same repository.