In this benchmark run

LoraDB is fastest overall across 82 workloads

Geometric-mean slowdown per engine across every workload they share. LoraDB wins eight of twelve groups; Grafeo takes setup and writes; Kuzu takes strings and numerics. Numbers here come straight from comparisons/report.md — same suite, same fixtures, same seed.

82 workloads12 groups6 comparison engines

1×

LoraDB total

Fastest overall across this suite.

Workloads

12 groups · 6 engines compared

8/12

Groups won by LoraDB

Grafeo wins 2; Kuzu wins 2.

2.30×

vs. next-best total

Next-best (Grafeo) 2.30× slower overall.

Dive into a head-to-head — each page compares only LoraDB against that engine.

Where LoraDB leads

Strongest across scans, predicates, traversals, and patterns.

LoraDB wins the group on every workload it touches in these areas. Strings and numerics belong to Kuzu (LoraDB is within ~22% on both); writes and setup belong to Grafeo (LoraDB pays roughly 2× on writes against Grafeo’s purpose-built crate path).

scans6 workloads
predicates12 workloads
aggregates9 workloads
pipeline9 workloads
lists3 workloads
sort3 workloads
traversals15 workloads
patterns4 workloads

Overall ranking

Total geometric-mean slowdown, ordered fastest to slowest.

LoraDBfastest

Grafeo2.30× slower

Kuzu3.30× slower

Neo4j7.99× slower

Memgraph9.75× slower

SurrealDB48.6× slower

HelixDB57.2× slower

Log scale. Bar length reflects log₁₀(slowdown) so multiples spanning 2.3×–57× remain visually comparable. Slowdowns are copied verbatim from comparisons/report.md.

Per group

Twelve groups, eighty-two workloads.

Each card shows the workload count, the group winner, and the per-engine slowdown chip. LoraDB sits in its own row so the comparison stays anchored even when LoraDB is not the row winner.

setup

1 workload

WinnerGrafeo

LoraDB2.70× slower

Kuzu700×
Grafeofastest
SurrealDBn/a
Memgraph904×
Neo4j1769×
HelixDB1247×

writes

9 workloads

WinnerGrafeo

LoraDB0.61× slower

Kuzu1.79×
Grafeofastest
SurrealDB28.3×
Memgraph3.71×
Neo4j7.75×
HelixDB125×

scans

6 workloads

WinnerLoraDB

LoraDBfastest

Kuzu6.37×
Grafeo5.43×
SurrealDB121×
Memgraph27.5×
Neo4j24.1×
HelixDB287×

predicates

12 workloads

WinnerLoraDB

LoraDBfastest

Kuzu1.14×
Grafeo1.83×
SurrealDB21.3×
Memgraph4.77×
Neo4j3.78×
HelixDB20.9×

strings

5 workloads

WinnerKuzu

LoraDB1.22× slower

Kuzufastest
Grafeo1.84×
SurrealDB34.1×
Memgraph11.3×
Neo4j4.19×
HelixDBn/a

numerics

6 workloads

WinnerKuzu

LoraDB1.08× slower

Kuzufastest
Grafeo1.87×
SurrealDB39.9×
Memgraph10.7×
Neo4j4.00×
HelixDB20.4×

aggregates

9 workloads

WinnerLoraDB

LoraDBfastest

Kuzu3.18×
Grafeo1.89×
SurrealDB44.7×
Memgraph6.25×
Neo4j6.94×
HelixDB35.5×

pipeline

9 workloads

WinnerLoraDB

LoraDBfastest

Kuzu1.27×
Grafeo1.30×
SurrealDB36.8×
Memgraph4.97×
Neo4j3.07×
HelixDB17.2×

lists

3 workloads

WinnerLoraDB

LoraDBfastest

Kuzu11.3×
Grafeo3.21×
SurrealDB32.3×
Memgraph41.6×
Neo4j42.2×
HelixDB15.0×

sort

3 workloads

WinnerLoraDB

LoraDBfastest

Kuzu1.41×
Grafeo1.60×
SurrealDB31.8×
Memgraph4.28×
Neo4j3.48×
HelixDB16.4×

traversals

15 workloads

WinnerLoraDB

LoraDBfastest

Kuzu20.4×
Grafeo5.88×
SurrealDB110×
Memgraph22.4×
Neo4j17.8×
HelixDB133×

patterns

4 workloads

WinnerLoraDB

LoraDBfastest

Kuzu2.80×
Grafeo2.45×
SurrealDB190×
Memgraph7.59×
Neo4j4.96×
HelixDB22.2×

Summary table

The summary in one table.

Each engine column carries the geometric-mean slowdown of that engine vs the group winner across every workload they share. Empty cells are noted below.

Geometric-mean slowdown by group, lower is better.
Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
setup	1	2.70× slower	700× slower	fastest	omitted	904× slower	1769× slower	1247× slower	Grafeo
writes	9	0.61× slower	1.79× slower	fastest	28.3× slower	3.71× slower	7.75× slower	125× slower	Grafeo
scans	6	fastest	6.37× slower	5.43× slower	121× slower	27.5× slower	24.1× slower	287× slower	LoraDB
predicates	12	fastest	1.14× slower	1.83× slower	21.3× slower	4.77× slower	3.78× slower	20.9× slower	LoraDB
strings	5	1.22× slower	fastest	1.84× slower	34.1× slower	11.3× slower	4.19× slower	omitted	Kuzu
numerics	6	1.08× slower	fastest	1.87× slower	39.9× slower	10.7× slower	4.00× slower	20.4× slower	Kuzu
aggregates	9	fastest	3.18× slower	1.89× slower	44.7× slower	6.25× slower	6.94× slower	35.5× slower	LoraDB
pipeline	9	fastest	1.27× slower	1.30× slower	36.8× slower	4.97× slower	3.07× slower	17.2× slower	LoraDB
lists	3	fastest	11.3× slower	3.21× slower	32.3× slower	41.6× slower	42.2× slower	15.0× slower	LoraDB
sort	3	fastest	1.41× slower	1.60× slower	31.8× slower	4.28× slower	3.48× slower	16.4× slower	LoraDB
traversals	15	fastest	20.4× slower	5.88× slower	110× slower	22.4× slower	17.8× slower	133× slower	LoraDB
patterns	4	fastest	2.80× slower	2.45× slower	190× slower	7.59× slower	4.96× slower	22.2× slower	LoraDB
total	82	fastest	3.30× slower	2.30× slower	48.6× slower	9.75× slower	7.99× slower	57.2× slower	LoraDB

Per-workload detail

Every workload, every engine, raw timings.

Workload rows show the raw mean time plus the slowdown against the row’s winner. Omitted cells mean the workload has no like-for-like equivalent on that engine; the reason is captured under each group.

setup(1)

WinnerGrafeo

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
construct_empty	—	4.06 µs2.70× slower	1.05 ms700× slower	1.50 µsfastest	omitted	1.36 ms904× slower	2.66 ms1769× slower	1.87 ms1247× slower	Grafeo

writes(9)

WinnerGrafeo

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
bulk_edges	200	606.63 µsfastest	2.01 ms3.31× slower	19.86 ms32.7× slower	omitted	3.18 ms5.24× slower	2.65 ms4.37× slower	309.68 ms510× slower	LoraDB
bulk_set_match	1000	552.39 µs1.84× slower	300.03 µsfastest	315.88 µs1.05× slower	6.15 ms20.5× slower	726.92 µs2.42× slower	916.11 µs3.05× slower	3.57 ms11.9× slower	Kuzu
delete_node	1000	355.70 µs1.47× slower	598.35 µs2.48× slower	241.43 µsfastest	5.56 ms23.0× slower	483.58 µs2.00× slower	1.65 ms6.84× slower	151.74 ms629× slower	Grafeo
merge_create	1000	112.57 µs1.25× slower	670.68 µs7.45× slower	90.02 µsfastest	omitted	677.96 µs7.53× slower	2.48 ms27.6× slower	omitted	Grafeo
merge_existing	1000	23.95 µs2.89× slower	162.06 µs19.6× slower	8.29 µsfastest	omitted	304.94 µs36.8× slower	647.72 µs78.1× slower	omitted	Grafeo
set_multiple_props	1000	21.15 µsfastest	169.72 µs8.03× slower	182.77 µs8.64× slower	5.54 ms262× slower	313.89 µs14.8× slower	571.61 µs27.0× slower	2.37 ms112× slower	LoraDB
update_set	1000	18.01 µsfastest	188.62 µs10.5× slower	208.61 µs11.6× slower	6.04 ms335× slower	315.49 µs17.5× slower	732.47 µs40.7× slower	2.91 ms162× slower	LoraDB
write_bulk	1000	1.46 ms1.93× slower	2.76 ms3.65× slower	756.21 µsfastest	26.66 ms35.3× slower	3.13 ms4.14× slower	6.34 ms8.38× slower	771.22 ms1020× slower	Grafeo
write_single	1000	14.90 µs2.20× slower	omitted	6.78 µsfastest	252.56 µs37.3× slower	413.52 µs61.0× slower	1.30 ms192× slower	153.86 ms22704× slower	Grafeo

Omissions

Workloads omitted only where the engine has no like-for-like equivalent — a missing scalar function, a different MERGE semantics, a reserved word. The reason is recorded with the workload, not silently dropped.

write_single
Kuzu
- Kuzu requires `CREATE NODE TABLE` before inserts; the empty fixture has no schema and adding one would change the iteration's measured cost.
merge_existing
HelixDBSurrealDB
- HelixDB has no MERGE/upsert; emulating it needs conditional var_as_if branching.
- SurrealDB's UPSERT semantics diverge from Cypher MERGE on which fields are matched vs set.
merge_create
HelixDBSurrealDB
- HelixDB has no MERGE/upsert (see merge_existing).
- See merge_existing — UPSERT semantics differ.
bulk_edges
SurrealDB
- UNWIND-driven bulk RELATE requires scripted FOR loops; not a like-for-like comparison.

scans(6)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
distinct	1000	198.38 µsfastest	448.06 µs2.26× slower	246.37 µs1.24× slower	omitted	738.04 µs3.72× slower	566.24 µs2.85× slower	omitted	LoraDB
lookup_by_id	1000	716.64 nsfastest	116.07 µs162× slower	170.03 µs237× slower	5.26 ms7340× slower	305.39 µs426× slower	601.36 µs839× slower	2.30 ms3205× slower	LoraDB
lookup_by_id_indexed	1000	684.00 nsfastest	116.18 µs170× slower	22.73 µs33.2× slower	32.35 µs47.3× slower	310.02 µs453× slower	747.01 µs1092× slower	2.62 ms3831× slower	LoraDB
range_filter	1000	199.70 µs1.04× slower	192.68 µsfastest	214.22 µs1.11× slower	9.29 ms48.2× slower	1.22 ms6.31× slower	589.16 µs3.06× slower	omitted	Kuzu
scan_filtered	1000	149.08 µsfastest	158.49 µs1.06× slower	211.95 µs1.42× slower	5.88 ms39.5× slower	1.15 ms7.73× slower	718.45 µs4.82× slower	3.82 ms25.6× slower	LoraDB
scan_label	1000	124.96 µsfastest	131.34 µs1.05× slower	213.71 µs1.71× slower	5.01 ms40.1× slower	1.61 ms12.9× slower	661.84 µs5.30× slower	2.70 ms21.6× slower	LoraDB

Omissions

distinct
HelixDBSurrealDB
- HelixDB's dedup is node-level; there is no SELECT DISTINCT <property>.
- `value` is a reserved word in SurrealQL's SELECT VALUE clause; no clean equivalent.

predicates(12)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
where_compound_and_or	1000	213.07 µsfastest	230.83 µs1.08× slower	646.50 µs3.03× slower	10.23 ms48.0× slower	997.22 µs4.68× slower	584.96 µs2.75× slower	3.23 ms15.2× slower	LoraDB
where_contains	1000	145.04 µsfastest	166.24 µs1.15× slower	255.53 µs1.76× slower	4.96 ms34.2× slower	638.95 µs4.41× slower	597.31 µs4.12× slower	3.02 ms20.8× slower	LoraDB
where_ends_with	1000	143.60 µsfastest	167.41 µs1.17× slower	236.22 µs1.64× slower	5.06 ms35.2× slower	726.66 µs5.06× slower	581.26 µs4.05× slower	2.99 ms20.8× slower	LoraDB
where_id_in_range	1000	142.71 µs2.02× slower	187.97 µs2.66× slower	70.55 µsfastest	7.54 ms107× slower	433.95 µs6.15× slower	663.12 µs9.40× slower	3.13 ms44.3× slower	Grafeo
where_in_list	1000	162.42 µsfastest	196.06 µs1.21× slower	270.30 µs1.66× slower	5.59 ms34.4× slower	606.63 µs3.74× slower	627.53 µs3.86× slower	2.92 ms18.0× slower	LoraDB
where_modulo_eq	1000	127.32 µsfastest	167.83 µs1.32× slower	250.88 µs1.97× slower	128.48 µs1.01× slower	717.29 µs5.63× slower	557.09 µs4.38× slower	3.01 ms23.7× slower	LoraDB
where_not	1000	166.05 µs1.01× slower	164.07 µsfastest	340.52 µs2.08× slower	8.27 ms50.4× slower	1.24 ms7.57× slower	695.07 µs4.24× slower	4.64 ms28.3× slower	Kuzu
where_or	1000	147.16 µsfastest	185.19 µs1.26× slower	450.08 µs3.06× slower	8.43 ms57.3× slower	627.82 µs4.27× slower	603.10 µs4.10× slower	3.29 ms22.3× slower	LoraDB
where_starts_with	1000	146.91 µsfastest	168.08 µs1.14× slower	249.22 µs1.70× slower	5.12 ms34.9× slower	727.06 µs4.95× slower	596.02 µs4.06× slower	3.29 ms22.4× slower	LoraDB
where_string_gte	1000	180.97 µs1.07× slower	168.40 µsfastest	247.34 µs1.47× slower	6.05 ms35.9× slower	1.20 ms7.10× slower	612.78 µs3.64× slower	4.11 ms24.4× slower	Kuzu
where_subexpr	1000	227.59 µs1.69× slower	194.85 µs1.45× slower	589.01 µs4.37× slower	134.71 µsfastest	1.76 ms13.1× slower	584.16 µs4.34× slower	4.73 ms35.1× slower	SurrealDB
where_two_props	1000	152.69 µsfastest	205.41 µs1.35× slower	404.15 µs2.65× slower	6.46 ms42.3× slower	390.30 µs2.56× slower	589.64 µs3.86× slower	2.56 ms16.8× slower	LoraDB

strings(5)

WinnerKuzu

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
string_concat	1000	186.55 µs1.27× slower	147.24 µsfastest	283.23 µs1.92× slower	5.86 ms39.8× slower	1.68 ms11.4× slower	577.43 µs3.92× slower	omitted	Kuzu
string_size	1000	172.17 µs1.10× slower	156.38 µsfastest	251.12 µs1.61× slower	5.23 ms33.5× slower	1.70 ms10.9× slower	567.80 µs3.63× slower	omitted	Kuzu
string_substring	1000	210.20 µs1.21× slower	173.57 µsfastest	314.23 µs1.81× slower	5.58 ms32.1× slower	2.06 ms11.9× slower	877.36 µs5.05× slower	omitted	Kuzu
string_to_lower	1000	201.00 µs1.33× slower	151.22 µsfastest	297.66 µs1.97× slower	5.03 ms33.3× slower	1.72 ms11.4× slower	684.21 µs4.52× slower	omitted	Kuzu
string_to_upper	1000	185.41 µs1.22× slower	151.96 µsfastest	292.95 µs1.93× slower	4.91 ms32.3× slower	1.67 ms11.0× slower	603.96 µs3.97× slower	omitted	Kuzu

Omissions

string_to_upper
HelixDB
- HelixDB's DSL has no scalar string functions (upper/lower/substring/length/concat) — only property filters and graph traversal.
string_to_lower
HelixDB
- No scalar string functions in the DSL (see string_to_upper).
string_substring
HelixDB
- No scalar string functions in the DSL (see string_to_upper).
string_size
HelixDB
- No scalar string functions in the DSL (see string_to_upper).
string_concat
HelixDB
- No scalar string functions in the DSL (see string_to_upper).

numerics(6)

WinnerKuzu

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
numeric_abs	1000	174.85 µs1.13× slower	154.99 µsfastest	272.48 µs1.76× slower	5.86 ms37.8× slower	1.64 ms10.6× slower	636.05 µs4.10× slower	omitted	Kuzu
numeric_ceil	1000	171.69 µs1.10× slower	155.40 µsfastest	269.13 µs1.73× slower	5.81 ms37.4× slower	1.78 ms11.5× slower	634.65 µs4.08× slower	omitted	Kuzu
numeric_floor	1000	175.48 µs1.10× slower	159.62 µsfastest	326.82 µs2.05× slower	5.99 ms37.5× slower	1.64 ms10.3× slower	609.30 µs3.82× slower	omitted	Kuzu
numeric_modulo	1000	144.95 µs1.02× slower	142.67 µsfastest	236.39 µs1.66× slower	omitted	1.59 ms11.2× slower	637.61 µs4.47× slower	2.97 ms20.8× slower	Kuzu
numeric_pow	1000	166.22 µs1.16× slower	143.34 µsfastest	394.61 µs2.75× slower	8.42 ms58.8× slower	1.68 ms11.7× slower	641.94 µs4.48× slower	2.87 ms20.0× slower	Kuzu
numeric_round	1000	180.68 µsfastest	181.69 µs1.01× slower	269.75 µs1.49× slower	5.87 ms32.5× slower	1.68 ms9.30× slower	581.78 µs3.22× slower	omitted	LoraDB

Omissions

numeric_abs
HelixDB
- HelixDB's Expr supports +,-,*,/,% but no abs/floor/ceil/round.
numeric_modulo
SurrealDB
- SurrealQL parser rejects bare `%` inside SELECT projections (same parse limit as grouped_aggregation).
numeric_floor
HelixDB
- HelixDB's Expr supports +,-,*,/,% but no abs/floor/ceil/round.
numeric_ceil
HelixDB
- HelixDB's Expr supports +,-,*,/,% but no abs/floor/ceil/round.
numeric_round
HelixDB
- HelixDB's Expr supports +,-,*,/,% but no abs/floor/ceil/round.

aggregates(9)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
aggregate_avg	1000	81.43 µsfastest	253.54 µs3.11× slower	202.26 µs2.48× slower	6.22 ms76.4× slower	562.93 µs6.91× slower	606.76 µs7.45× slower	3.20 ms39.3× slower	LoraDB
aggregate_collect	1000	78.90 µsfastest	275.13 µs3.49× slower	214.58 µs2.72× slower	6.79 ms86.0× slower	599.26 µs7.60× slower	619.24 µs7.85× slower	omitted	LoraDB
aggregate_count	1000	59.50 µs2.77× slower	250.00 µs11.7× slower	21.44 µsfastest	209.05 µs9.75× slower	381.02 µs17.8× slower	588.34 µs27.4× slower	omitted	Grafeo
aggregate_count_distinct	1000	104.05 µsfastest	441.70 µs4.24× slower	213.10 µs2.05× slower	omitted	540.27 µs5.19× slower	635.68 µs6.11× slower	omitted	LoraDB
aggregate_max	1000	79.97 µsfastest	263.33 µs3.29× slower	200.95 µs2.51× slower	6.08 ms76.0× slower	572.73 µs7.16× slower	619.56 µs7.75× slower	2.86 ms35.7× slower	LoraDB
aggregate_min	1000	79.67 µsfastest	258.01 µs3.24× slower	197.17 µs2.47× slower	5.95 ms74.7× slower	552.13 µs6.93× slower	563.60 µs7.07× slower	2.49 ms31.3× slower	LoraDB
aggregate_sum	1000	78.09 µsfastest	263.06 µs3.37× slower	199.65 µs2.56× slower	6.23 ms79.8× slower	524.56 µs6.72× slower	568.81 µs7.28× slower	2.82 ms36.1× slower	LoraDB
grouped_aggregation	1000	154.46 µsfastest	549.19 µs3.56× slower	269.45 µs1.74× slower	omitted	751.51 µs4.87× slower	969.01 µs6.27× slower	omitted	LoraDB
top_k	1000	186.50 µsfastest	248.58 µs1.33× slower	424.61 µs2.28× slower	6.40 ms34.3× slower	955.44 µs5.12× slower	792.15 µs4.25× slower	omitted	LoraDB

Omissions

aggregate_count
HelixDB
- The HelixDB enterprise-dev image rejects `count()` dynamic queries with `rate limit exceeded` (its other 9 workloads run fine). The count handler is wired in helixdb.rs and would run on a server without that limit.
aggregate_collect
HelixDB
- aggregate_by offers Count/Sum/Min/Max/Mean, no list collect.
aggregate_count_distinct
HelixDBSurrealDB
- Needs count(DISTINCT); count() is rate-limited on the enterprise-dev image and value-level DISTINCT isn't exposed.
- count(DISTINCT) has no direct SurrealQL aggregate; requires nested SELECT + array::distinct.
grouped_aggregation
HelixDBSurrealDB
- group_count groups by a stored property, not a computed value % 10 key.
- SurrealQL rejects `%` inside a SELECT projection that's then used as a GROUP BY key (parse error).

pipeline(9)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
case_when	1000	173.33 µsfastest	186.25 µs1.07× slower	256.32 µs1.48× slower	7.38 ms42.6× slower	1.86 ms10.8× slower	796.60 µs4.60× slower	2.78 ms16.1× slower	LoraDB
coalesce_existing	1000	161.92 µsfastest	162.41 µs1.00× slower	260.13 µs1.61× slower	7.51 ms46.4× slower	1.62 ms9.99× slower	628.36 µs3.88× slower	3.32 ms20.5× slower	LoraDB
computed_in_return	1000	151.15 µs1.07× slower	140.92 µsfastest	241.22 µs1.71× slower	7.35 ms52.1× slower	1.61 ms11.5× slower	639.09 µs4.54× slower	2.99 ms21.2× slower	Kuzu
distinct_with_order	1000	508.01 µs2.08× slower	549.89 µs2.25× slower	243.88 µsfastest	omitted	692.42 µs2.84× slower	617.76 µs2.53× slower	omitted	Grafeo
predicate_via_function	1000	238.39 µs1.38× slower	172.38 µsfastest	434.70 µs2.52× slower	7.45 ms43.2× slower	1.77 ms10.3× slower	544.38 µs3.16× slower	omitted	Kuzu
with_aggregate_then_filter	1000	148.70 µsfastest	475.77 µs3.20× slower	267.49 µs1.80× slower	omitted	567.44 µs3.82× slower	710.00 µs4.77× slower	omitted	LoraDB
with_distinct_then_count	1000	202.72 µsfastest	540.85 µs2.67× slower	244.75 µs1.21× slower	omitted	597.19 µs2.95× slower	812.09 µs4.01× slower	omitted	LoraDB
with_pipeline	1000	186.41 µsfastest	333.60 µs1.79× slower	218.95 µs1.17× slower	5.98 ms32.1× slower	678.15 µs3.64× slower	581.26 µs3.12× slower	omitted	LoraDB
with_two_chained	1000	313.64 µs1.41× slower	222.76 µsfastest	388.95 µs1.75× slower	8.12 ms36.5× slower	1.22 ms5.46× slower	608.48 µs2.73× slower	4.25 ms19.1× slower	Kuzu

Omissions

with_pipeline
HelixDB
- Returns count(...); count() is rate-limited on the enterprise-dev image (see aggregate_count).
with_distinct_then_count
HelixDBSurrealDB
- count() rate-limited and no value-level DISTINCT.
- DISTINCT on `value` needs SELECT VALUE, where `value` is reserved (same parse limit as the `distinct` workload).
with_aggregate_then_filter
HelixDBSurrealDB
- Group-then-having on a computed key (value % 10) isn't expressible; group_count keys on a stored property.
- Groups on `value % 10`; SurrealQL rejects `%` in a projection used as a GROUP BY key (same parse limit as grouped_aggregation).
predicate_via_function
HelixDB
- WHERE size(name) needs a length() scalar the DSL doesn't provide.
distinct_with_order
HelixDBSurrealDB
- No value-level DISTINCT (see distinct).
- DISTINCT + ORDER BY on `value`, which is reserved in both SELECT VALUE and ORDER BY (see distinct, order_by_multi_key).

lists(3)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
list_in_construction	1000	177.94 µs1.05× slower	168.89 µsfastest	450.42 µs2.67× slower	6.14 ms36.4× slower	2.05 ms12.1× slower	733.38 µs4.34× slower	2.67 ms15.8× slower	Kuzu
list_unwind_explicit	1000	1.11 µsfastest	165.87 µs149× slower	7.11 µs6.40× slower	33.52 µs30.2× slower	322.60 µs291× slower	689.53 µs621× slower	omitted	LoraDB
range_function	1000	18.86 µsfastest	193.49 µs10.3× slower	38.62 µs2.05× slower	omitted	404.85 µs21.5× slower	555.15 µs29.4× slower	omitted	LoraDB

Omissions

range_function
SurrealDB
- SurrealQL has no row-generating numeric range (no UNWIND/range equivalent); the explicit-list unwind is covered by list_unwind_explicit.

sort(3)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
order_by_id_asc	1000	167.95 µsfastest	234.56 µs1.40× slower	226.74 µs1.35× slower	5.24 ms31.2× slower	738.63 µs4.40× slower	718.84 µs4.28× slower	3.20 ms19.1× slower	LoraDB
order_by_multi_key	1000	211.99 µsfastest	276.69 µs1.31× slower	442.71 µs2.09× slower	omitted	871.82 µs4.11× slower	554.84 µs2.62× slower	2.80 ms13.2× slower	LoraDB
skip_limit	1000	162.19 µsfastest	251.28 µs1.55× slower	233.92 µs1.44× slower	5.24 ms32.3× slower	704.66 µs4.34× slower	612.55 µs3.78× slower	2.87 ms17.7× slower	LoraDB

Omissions

order_by_multi_key
SurrealDB
- `value` is reserved in SurrealQL ORDER BY clauses (parse error).

traversals(15)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
direct_record_traversal	500	831.70 nsfastest	255.86 µs308× slower	64.09 µs77.1× slower	62.98 µs75.7× slower	384.41 µs462× slower	616.93 µs742× slower	1.55 ms1861× slower	LoraDB
recursive_depth2	500	968.69 nsfastest	499.89 µs516× slower	62.29 µs64.3× slower	110.22 µs114× slower	380.49 µs393× slower	601.14 µs621× slower	1.69 ms1742× slower	LoraDB
recursive_depth3	500	1.04 µsfastest	540.33 µs520× slower	62.87 µs60.5× slower	147.78 µs142× slower	382.45 µs368× slower	592.76 µs570× slower	1.46 ms1402× slower	LoraDB
recursive_depth5	500	1.16 µsfastest	511.05 µs440× slower	62.30 µs53.7× slower	225.94 µs195× slower	426.52 µs367× slower	606.31 µs522× slower	1.65 ms1425× slower	LoraDB
relation_filter	500	117.78 µsfastest	406.36 µs3.45× slower	omitted	18.70 ms159× slower	868.53 µs7.37× slower	586.82 µs4.98× slower	2.50 ms21.2× slower	LoraDB
traversal_count_one_hop	500	59.36 µsfastest	288.00 µs4.85× slower	141.39 µs2.38× slower	106.91 µs1.80× slower	394.49 µs6.65× slower	609.51 µs10.3× slower	omitted	LoraDB
traversal_filter_one_hop	500	138.99 µsfastest	405.25 µs2.92× slower	259.07 µs1.86× slower	21.51 ms155× slower	1.04 ms7.49× slower	582.58 µs4.19× slower	3.15 ms22.7× slower	LoraDB
traversal_one_hop	500	125.98 µsfastest	360.18 µs2.86× slower	263.65 µs2.09× slower	25.78 ms205× slower	1.14 ms9.06× slower	570.56 µs4.53× slower	2.59 ms20.5× slower	LoraDB
traversal_reverse	500	123.67 µsfastest	365.12 µs2.95× slower	266.69 µs2.16× slower	24.71 ms200× slower	1.10 ms8.93× slower	591.25 µs4.78× slower	2.62 ms21.2× slower	LoraDB
traversal_three_hop	500	236.10 µsfastest	1.13 ms4.80× slower	556.10 µs2.36× slower	60.59 ms257× slower	1.23 ms5.21× slower	607.69 µs2.57× slower	5.16 ms21.9× slower	LoraDB
traversal_two_hop	500	165.58 µsfastest	649.69 µs3.92× slower	411.83 µs2.49× slower	44.18 ms267× slower	1.27 ms7.66× slower	555.27 µs3.35× slower	64.61 ms390× slower	LoraDB
traversal_undirected	500	213.41 µsfastest	492.51 µs2.31× slower	494.10 µs2.32× slower	omitted	1.76 ms8.26× slower	637.72 µs2.99× slower	4.20 ms19.7× slower	LoraDB
variable_length_path	100	86.24 µsfastest	2.64 ms30.6× slower	210.43 µs2.44× slower	omitted	809.87 µs9.39× slower	605.82 µs7.02× slower	omitted	LoraDB
varlen_2_to_5	100	123.74 µsfastest	3.90 ms31.5× slower	274.33 µs2.22× slower	omitted	986.89 µs7.98× slower	578.29 µs4.67× slower	omitted	LoraDB
varlen_exact_5	100	56.89 µsfastest	3.86 ms67.9× slower	140.52 µs2.47× slower	omitted	576.13 µs10.1× slower	586.35 µs10.3× slower	omitted	LoraDB

Omissions

traversal_undirected
SurrealDB
- SurrealDB graph edges are directional; the undirected `-[:NEXT]-` pattern has no single like-for-like arrow form (forward and reverse are covered by traversal_one_hop and traversal_reverse).
variable_length_path
HelixDBSurrealDB
- Range-bounded variable-length expansion isn't mapped; fixed depths are benched as recursive_depth2/3/5.
- SurrealQL recursive traversal takes a fixed depth `@{n}` (see recursive_depth2/3/5); a `1..3` range expanded from every start node has no like-for-like form.
varlen_2_to_5
HelixDBSurrealDB
- Range-bounded variable-length expansion isn't mapped (see variable_length_path).
- Range-bounded variable-length expansion from every start node; see variable_length_path.
varlen_exact_5
HelixDBSurrealDB
- Range-bounded variable-length expansion isn't mapped (see variable_length_path).
- Fixed-depth expansion from every start node; the anchored single-source equivalent is benched as recursive_depth5.
relation_filter
Grafeo
- grafeo's `create_edge` facade takes no edge properties, so the chain fixture has no `step` to filter on. memgraph/neo4j/kuzu seed `step` directly, so they do run this workload.

patterns(4)

WinnerLoraDB

Workload	Size	LoraDB	Kuzu	Grafeo	SurrealDB	Memgraph	Neo4j	HelixDB	Winner
edge_subquery_clause	500	213.55 µsfastest	465.09 µs2.18× slower	omitted	18.19 ms85.2× slower	1.18 ms5.52× slower	605.92 µs2.84× slower	4.12 ms19.3× slower	LoraDB
star_fanout	1000	138.96 µsfastest	296.68 µs2.13× slower	321.31 µs2.31× slower	29.94 ms215× slower	1.62 ms11.7× slower	579.41 µs4.17× slower	2.75 ms19.8× slower	LoraDB
star_fanout_count	1000	61.36 µsfastest	290.42 µs4.73× slower	193.38 µs3.15× slower	20.35 ms332× slower	500.19 µs8.15× slower	590.08 µs9.62× slower	omitted	LoraDB
star_fanout_filter	1000	112.53 µsfastest	312.37 µs2.78× slower	227.60 µs2.02× slower	24.13 ms214× slower	707.87 µs6.29× slower	599.26 µs5.33× slower	3.24 ms28.8× slower	LoraDB

Omissions

star_fanout_count
HelixDB
- The HelixDB enterprise-dev image rejects `count()` dynamic queries with `rate limit exceeded` (its other 9 workloads run fine); see aggregate_count.
edge_subquery_clause
Grafeo
- grafeo's `create_edge` facade takes no edge properties, so the social fixture has no `strength` to filter on. memgraph/neo4j/kuzu seed `strength` directly, so they do run this workload.

Methodology

What this benchmark does and doesn’t say.

What’s measured

Mean iteration time on identical fixtures. Engines run the same query semantics where the language allows expressing them; where it doesn’t, the workload is marked omitted with the reason recorded next to it. Geometric mean is used at the group level so a single outlier doesn’t dominate the row.

What this isn’t

Not a production-scale TPC benchmark. Not a load test. Not a comparison of operational footprint or durability — every engine is run in-process or against a local server with the same warm fixture each iteration. Workloads are micro-benchmarks of the query pipeline, not of full systems.

Where to dig in

The full source — fixtures, harness, per-engine glue — lives in comparisons/. The report itself is generated from a single Criterion run; re-running the suite regenerates report.md, which this page is derived from.