Windowed High Cardinality
The low-cardinality bench's tumbling 30s window, aggregating count + avg(latency_ms) grouped by user_id from a key pool PINNED across products (500k users — see tenants below) — one open window per active user. Stresses the keyed-state store rather than framework overhead. Capacity is read in drain mode, on a backlog carrying synthetic event time (every product meets the same windows and keys); the live runs, at each product's max sustainable single-partition rate (its knee), give the close latency.
Last updated: 2026-11-10 · same-machine, relative comparison.
flowchart LR
src[("bench-input<br/>1 partition")]
s0{{"<b>key by</b><br/>500k keys (user_id)"}}
s1[["<b>window</b><br/>tumbling 30 s, event time<br/>count, avg(latency_ms)"]]
sink[("bench-out")]
state[("keyed state<br/>one open window per key")]
probe>"latency probe<br/>arrival − window_end_ms"]
src --> s0
s0 --> s1
s1 -->|"at-least-once"| sink
s1 -.- state
sink -.-> probeBench characteristics
In-process computation — Tumbling 30s event-time window on the Kafka message timestamp, 1 s grace, GROUP BY user_id (key pool pinned to 500k users — identical for every product regardless of its rate) → count + avg(latency_ms); one pane per key per window close.
| product | delivery | processing | engine | scaling on one partition | decode | state |
|---|---|---|---|---|---|---|
| turbine | at-least-once (idempotent, acks=all) | batch (Arrow columnar; 100k rows / 500 ms) | in-process · Rust core + free-threaded CPython 3.14t | intra-partition shards (parallelism=N, hash of partition_key, in-process) |
schema projection (6 of 19 fields, Arrow) | own RocksDB store (merge operator for windows) |
| turbine-resident | at-least-once (idempotent, acks=all) | batch (Arrow columnar; 100k rows / 500 ms) | in-process · Rust core + free-threaded CPython 3.14t | none — one unkeyed shard per partition (resident storage refuses parallelism) |
schema projection (6 of 19 fields, Arrow) | resident typed columns in memory, checkpointed whole (storage="resident") |
| arroyo | at-least-once (checkpointed) | batch (vectorized SQL, Arrow / DataFusion; source batches of 100k rows / 500 ms) | in-process · Rust (arroyo run: controller + one worker process) |
hash shuffle to N subtasks (arroyo run --parallelism, in-process queues) |
schema projection (2 of 19 fields) | in-memory tables, checkpointed to the state dir |
| flink | at-least-once (checkpointed) | batch (pipelined, JVM code-gen SQL) | in-process · JVM (local MiniCluster; Python only submits the job) | keyed exchange to N slots (parallelism.default, in-process local channels; source stays 1 subtask) |
schema projection (3 of 19 fields, Jackson); connector-default fetch | RocksDB (embedded, on-disk LSM) |
| quix | at-least-once (idempotent, acks=all) | row (per-record StreamingDataFrame) | in-process · pure-Python over librdkafka (C client) | none (one consumer per partition, GIL) | full record (19 of 19 fields, orjson dict) | RocksDB (embedded); changelog topic OFF for the bench |
| risingwave | at-least-once | batch (vectorized SQL) | external server · Rust (single-node: meta+compute+frontend) | hash exchange to N stream actors (SET streaming_parallelism, in-process) |
schema projection (2 of 19 fields) | server-managed (Hummock LSM store) |
Input is a single Kafka partition (lz4). Each product may spread it over several cores through its own in-process mechanism (column scaling on one partition; what it actually used is cpu_avg); an in-process exchange is not a broker hop — the extra rec/rec column counts records that DID cross the broker on other topics. All products run at-least-once, windows are tumbling 30 s event-time on the Kafka message timestamp.
Consumer fetch: Turbine's core hard-codes librdkafka fetch.min.bytes=64 KB, fetch.wait.max.ms=10, max.partition.fetch.bytes=50 MB and a 500k-msg / ~2 GB prefetch queue (not adjustable), so the SAME client gets the SAME settings: quix and RisingWave (librdkafka) run with them, and so does FastStream (aiokafka, whose equivalent knobs lifted it ~14%). Flink runs at its Java connector's defaults: an A/B (2026-09-10) showed the identical numbers throttled the Java consumer (373k/s vs 543k/s at the same ~285k rows/s per core) — identical numbers are not identical behaviour on a different client. Memory at saturation is mostly this prefetch queue, for every librdkafka product.
Drain ceiling (--prefill: static backlog, no producer running)
| product | backlog | event time/s | warmup + window | drained/s | cpu_avg % | rows/s per core | rows/s per GB | µs CPU / record | panes/s | run |
|---|---|---|---|---|---|---|---|---|---|---|
| turbine 0.3.5a1 | 200M | 300k | 40 + 120 s | 458,202 | 294.2 | 155,737 | 146,435 | 6.42 | 24,714.40 | 2026-11-10 |
| turbine-resident 0.3.5a1 | 300M | 300k | 40 + 120 s | 858,879 | 287.2 | 299,018 | 444,738 | 3.34 | 45,257.50 | 2026-11-10 |
| arroyo 0.15.0 | 200M | 300k | 40 + 240 s | 596,249 | 123.7 | 481,907 | 431,098 | 2.08 | 33,019.85 | 2026-10-27 |
| flink 2.2.1 | 200M | 300k | 268 + 515 s | 57,064 | 178.6 | 31,959 | 32,612 | 31.29 | 3,111.83 | 2026-10-27 |
| quix 3.23.6 | 140M | 300k | 385 + 750 s | 33,191 | 103.3 | 32,143 | 19,557 | 31.11 | 1,993.86 | 2026-10-20 |
| risingwave v3.1.0 | 140M | 300k | 85 + 150 s | 107,751 | 206.1 | 52,286 | 7,946 | 19.13 | 6,622.10 | 2026-10-20 |
How to read this table
gendata wrote the backlog and exited BEFORE the consumer started, so the window measures a drain with no producer competing for the box — the product's consume ceiling free of gendata's CPU and 10 Hz burstiness. Not comparable with a live table (different regime); ⚠ exhausted = the backlog ran out mid-window, rate understated.
The backlog's records carry synthetic event time, event time/s records per second of it (gendata --event-time-rate): it stands for a live stream at that rate, so every product meets the SAME windows — the same records and keys per window, which live runs at each product's own knee do not — and drained/s is its capacity on that problem, closes included. Event time runs at drained ÷ event-time rate × the wall clock, panes/s with it. A product draining far below the event-time rate gets a longer warmup and window (≥ its first close, then two more, at its live knee's pace: warmup + window). No latency here: a pane's age is the backlog's. 0 ⚠ panes/s: no pane was published during the window — the drain did not run the whole pipeline, its drained/s is not comparable.
Throughput / efficiency (live, at each product's knee) — capped by the machine: gendata produces ~430k/s at most, so read capacity in the drain table
| product | offered/s | consumed/s | regime (lag drift) | cpu_avg % | rows/s per core | rows/s per GB | µs CPU / record | panes/s | extra rec/rec | run |
|---|---|---|---|---|---|---|---|---|---|---|
| turbine 0.3.5a1 | 422,132 | 418,445 | keep-up (+0.7%) | 231.8 | 180,546 | 135,006 | 5.54 | 16,587.46 | 0 | 2026-11-10 |
| turbine-resident 0.3.5a1 | 423,125 | 418,242 | keep-up (+1.1%) | 126.7 | 329,985 | 295,741 | 3.03 | 16,473.13 | 0 | 2026-11-10 |
| arroyo 0.15.0 | 395,609 | 393,786 | keep-up (+0.5%) | 86.3 | 456,105 | 392,350 | 2.19 | 16,547.97 | 0 | 2026-10-27 |
| flink 2.2.1 | 34,613 | 35,898 | keep-up (-3.6%) | 124.2 | 28,912 | 23,553 | 34.59 | 14,531.28 | 0 | 2026-10-27 |
| quix 3.23.6 | 23,746 | 23,410 | keep-up (+1.5%) | 91.9 | 25,475 | 21,696 | 39.25 | 12,562.70 | 0 | 2026-10-20 |
| risingwave v3.1.0 | 118,681 | 117,630 | keep-up (+0.9%) | 201.0 | 58,509 | 16,661 | 17.09 | 16,468.68 | 0 | 2026-10-20 |
How to read this table
consumed/s is what the product actually drained (its consumer-group offset or engine counter), offered/s what gendata produced; a dotted underline marks the median of a repeated batch — hover it for the run count and the min – max. regime (lag drift): keep-up = consumed ≈ offered with a flat backlog (a sustainable point); saturated = fed beyond capacity, so consumed/s IS the ceiling and latency is unbounded; draining = still eating warmup backlog (consumed > offered; CPU is catch-up, not steady state — lengthen the warmup, or raise the rate if the ceiling is what you want). The lag drift in parentheses = backlog change over the window ÷ records consumed — the evidence for the regime (≈ 0 at keep-up, positive when saturated, negative while draining). cpu_avg is the whole process tree, 100 = one core busy. rows/s per core = consumed ÷ (cpu_avg/100); rows/s per GB = consumed ÷ memory; extra rec/rec = records written to topics other than the input and output per record consumed (0 = no broker hop).
panes/s is the window-emission cadence (≈ keys ÷ window size), a liveness signal, not a throughput figure — it only moves when windows close, which needs the consumer to keep up.
Latency (close latency — pane published minus window end)
| product | regime | p50 ms | p90 ms | p99 ms | p99.9 ms | max ms | records | setup floor p50 / p99 ms | added p50 / p99 ms |
|---|---|---|---|---|---|---|---|---|---|
| turbine 0.3.5a1 | keep-up | 4,366.5 | 6,133.5 | 6,706.7 | 6,866.0 | 7,021.1 | 6058232 | 106.5 / 205.0 | 4,259.9 / 6,501.7 |
| turbine-resident 0.3.5a1 | keep-up | 3,367.6 | 4,651.3 | 5,493.2 | 5,666.1 | 5,745.0 | 5980024 | 106.5 / 204.8 | 3,261.1 / 5,288.4 |
| arroyo 0.15.0 | keep-up | 3,739.2 | 5,502.4 | 6,455.9 | 6,560.9 | 6,765.7 | 11958552 | 106.4 / 204.7 | 3,632.8 / 6,251.2 |
| flink 2.2.1 | keep-up | 2,994.8 | 11,183.5 | 15,776.1 | 16,273.1 | 16,508.1 | 10601271 | 106.5 / 204.7 | 2,888.3 / 15,571.5 |
| quix 3.23.6 | keep-up | 3,552.0 | 5,613.9 | 6,207.3 | 6,417.5 | 6,563.2 | 4547095 | 106.5 / 204.7 | 3,445.5 / 6,002.6 |
| risingwave v3.1.0 | keep-up | 5,286.0 | 5,778.8 | 5,980.9 | 6,068.3 | 6,273.0 | 5977776 | 106.5 / 204.7 | 5,179.5 / 5,776.2 |
How to read this table
Measured by a probe consumer tailing the output topic (C client, 10 ms max fetch wait — the probe adds ≲10 ms, the same for every product). Only shown for keep-up runs: under saturation the figure is the backlog's age and grows without bound. setup floor is a harness producer writing one small record every 10 ms (acks=all, 5 ms linger) to its own topic during the same run, measured by the same probe: the broker's produce→visible latency at the products' output shape (on Redpanda, small produces surface on its flush cadence), which every product inherits. added = product − floor, the part the product is responsible for (fetch, batching, commit/epoch cadence, produce). The floor is a low-volume topic's (100 small records/s, surfacing on Redpanda's 100 ms flush.ms): an output topic carrying several MB/s — every record scored and republished — surfaces sooner: a percentile under the floor's reads below, added does not apply. Colours: each percentile against the best (lowest) of its column — green under 2×, yellow from 2×, orange from 5×, red from 10× (the ratio on hover).
Resources (raw)
| product | cpu_avg % | cpu_p95 % | mem_avg MB | mem_max MB | samples | instrument |
|---|---|---|---|---|---|---|
| turbine 0.3.5a1 | 294.2 | 414.2 | 3,204.1 | 3,492.6 | 116 | /proc Δ, pss |
| turbine 0.3.5a1 | 231.8 | 354.7 | 3,173.9 | 3,425.3 | 116 | /proc Δ, pss |
| turbine-resident 0.3.5a1 | 126.7 | 146.9 | 1,448.2 | 1,504.7 | 117 | /proc Δ, pss |
| turbine-resident 0.3.5a1 | 287.2 | 303.9 | 1,977.6 | 2,041.1 | 116 | /proc Δ, pss |
| arroyo 0.15.0 | 123.7 | 191.8 | 1,416.3 | 1,672.1 | 234 | /proc Δ, pss |
| arroyo 0.15.0 | 86.3 | 109.4 | 1,027.7 | 1,235.8 | 235 | /proc Δ, pss |
| flink 2.2.1 | 178.6 | 348.1 | 1,791.8 | 1,839.7 | 504 | /proc Δ, pss |
| flink 2.2.1 | 124.2 | 366.9 | 1,560.7 | 1,641.6 | 234 | /proc Δ, pss |
| quix 3.23.6 | 91.9 | 116.4 | 1,104.9 | 1,226.6 | 117 | /proc Δ, pss |
| quix 3.23.6 | 103.3 | 118.3 | 1,737.9 | 1,968.2 | 730 | /proc Δ, pss |
| risingwave v3.1.0 | 206.1 | 272.2 | 13,886.1 | 16,870.8 | 147 | /proc Δ, rss |
| risingwave v3.1.0 | 201.0 | 331.4 | 7,229.6 | 13,487.0 | 118 | /proc Δ, rss |
How to read this table
CPU is summed over the whole process tree (JVM, UDF servers, the RisingWave container's processes) as utime+stime deltas per 1 s sample — 100 = one core. Memory is PSS (shared pages counted once) where readable, RSS otherwise (instrument). Rows marked ⚠ legacy ps were measured with the old ps %CPU lifetime average and are not comparable: re-run them.
Product notes
flink — Local MiniCluster (PyFlink only submits the job), parallelism = the run's cores (3 since 2026-10-20, 1 before), RocksDB keyed-state backend (EmbeddedRocksDBStateBackend, on-disk LSM) — apples-to-apples with quix/Turbine/RisingWave, NOT Flink's default heap (RAM) backend (which sustained 3.5× more at the PRE-PIN pool, but kept every window key in the JVM heap at ~3.5 GB). At-least-once checkpointing. CPU/RSS include the JVM child (see the stateless note). Under the pinned 500k-key pool the keyed-state close path (≈500k panes per close at any rate ≥ 60k) is the binding constraint, not ingestion — well under one core at the knee — so flink lands below quix, where the pre-pin pool had it above.
risingwave — Server-style product: the dataflow runs in the risingwave container (single node, streaming_parallelism = the run's cores: 3 since 2026-10-20, its adaptive default before — every core of the box), the app is a thin DDL client; CPU/RSS are the whole container, consumed/s is read from its source state (the Kafka offset its source has read). The container is re-created before each run and stopped after it: a clean state store, and no compaction left running into the next run. At the pinned 500k-key pool its keyed aggregation parallelises across actors and holds a rate an order of magnitude above the single-core in-process products — read its RSS with the whole-single-node-server caveat.
Sanity — ✓ matches theory
Are we benching the same thing?
| product | panes/s | expected | sample |
|---|---|---|---|
| turbine 0.3.5a1 · drain | 24,714.40 | 25,455.65 | 500 |
| turbine 0.3.5a1 | 16,587.46 | 16,666.67 | 500 |
| turbine-resident 0.3.5a1 | 16,473.13 | 16,666.67 | 500 |
| turbine-resident 0.3.5a1 · drain | 45,257.50 | 48,009.24 | 500 |
| arroyo 0.15.0 · drain | 33,019.85 | 33,124.92 | 500 |
| arroyo 0.15.0 | 16,547.97 | 16,666.67 | 500 |
| flink 2.2.1 · drain | 3,111.83 | 2,669.16 | 500 |
| flink 2.2.1 | 14,531.28 | 14,625.73 | 500 |
| quix 3.23.6 | 12,562.70 | 12,717.87 | 500 |
| quix 3.23.6 · drain | 1,993.86 | 1,843.94 | 500 |
| risingwave v3.1.0 · drain | 6,622.10 | 5,747.95 | 500 |
| risingwave v3.1.0 | 16,468.68 | 16,654.22 | 500 |
Aggregated output: checked against the expected pane cadence (expected.output_rate in bench.yaml; × drained ÷ event-time rate for a windowed drain), not per-record selectivity.
Verdict: ✓ matches theory — panes/s within 2× of the expected cadence for every row.
Turbine releases
| line | version | bench | regime | offered/s | consumed/s | rows/s per core | µs CPU / record | rows/s per GB | p99 ms | Δ efficiency | Δ capacity | run |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| turbine | 0.3.5a1 | 3 | keep-up | 422,132 | 418,445 | 180,546 | 5.54 | 135,006 | 6,706.7 | +2.1% | - | 2026-11-10 |
| turbine | 0.3.4a1 | 3 | keep-up | 421,384 | 417,602 | 176,847 | 5.65 | 120,480 | 6,922.4 | +6.6% | - | 2026-11-10 |
| turbine | 0.3.3a1 | 3 | keep-up | 425,059 | 417,302 | 165,853 | 6.03 | 117,920 | 11,918.6 | +10.6% | - | 2026-11-06 |
| turbine | 0.3.2a1 | 3 | keep-up | 423,918 | 409,818 | 149,967 | 6.67 | 107,166 | 37,406.5 | ≠ config | ≠ config | 2026-11-04 |
| turbine | 0.3.1a2 | 3 | keep-up | 385,768 | 383,376 | 154,814 | 6.46 | 107,239 | 35,021.2 | ≠ config | ≠ config | 2026-11-01 |
| turbine | 0.3.1a1 | 3 | keep-up | 296,864 | 294,616 | 141,761 | 7.05 | 91,054 | 31,615.3 | ≠ bench | ≠ bench | 2026-10-27 |
| turbine | 0.2.2a2 | 2 ⚠ deprecated | keep-up | 296,693 | 293,620 | 121,322 | 8.24 | 90,228 | 21,390.7 | +0.8% | - | 2026-10-20 |
| turbine | 0.2.2a1 | 2 ⚠ deprecated | keep-up | 296,807 | 293,736 | 120,304 | 8.31 | 91,082 | 21,623.0 | +4.4% | - | 2026-10-18 |
| turbine | 0.2.1a4 | 2 ⚠ deprecated | saturated | 296,770 | 276,846 | 115,249 | 8.68 | 79,834 | - | -2.7% | - | 2026-10-18 |
| turbine | 0.2.1a3 | 2 ⚠ deprecated | keep-up | 296,809 | 287,980 | 118,469 | 8.44 | 84,241 | 23,443.1 | +2.7% | - | 2026-10-18 |
| turbine | 0.2.1a2 | 2 ⚠ deprecated | saturated | 296,778 | 282,548 | 115,322 | 8.67 | 85,276 | - | -3.3% | +54.3% | 2026-10-18 |
| turbine | 0.2.0a2 | 2 ⚠ deprecated | saturated | 296,729 | 183,165 | 119,296 | 8.38 | 54,154 | - | - | - | 2026-10-18 |
| turbine · drain | 0.3.5a1 | 3 | drain | — | 458,202 | 155,737 | 6.42 | 146,435 | - | +5.8% | +11.4% | 2026-11-10 |
| turbine · drain | 0.3.4a1 | 3 | drain | — | 411,147 | 147,204 | 6.79 | 113,686 | - | -9.6% ⚠ | -10.3% ⚠ | 2026-11-10 |
| turbine · drain | 0.3.3a1 | 3 | drain | — | 458,159 | 162,812 | 6.14 | 119,340 | - | +9.5% | +20.7% | 2026-11-06 |
| turbine · drain | 0.3.2a1 | 3 | drain | — | 379,542 | 148,627 | 6.73 | 91,009 | - | -3.7% | -5.3% ⚠ | 2026-11-03 |
| turbine · drain | 0.3.1a2 | 3 | drain | — | 400,765 | 154,328 | 6.48 | 96,363 | - | +2.9% | +4.3% | 2026-11-01 |
| turbine · drain | 0.3.1a1 | 3 | drain | — | 384,177 | 149,945 | 6.67 | 91,627 | - | +2.9% | +0.5% | 2026-10-28 |
| turbine · drain | 0.2.2a2 | 3 | drain | — | 382,231 | 145,699 | 6.86 | 83,065 | - | - | - | 2026-10-27 |
| turbine-resident | 0.3.5a1 | 3 | keep-up | 423,125 | 418,242 | 329,985 | 3.03 | 295,741 | 5,493.2 | -2.4% | - | 2026-11-10 |
| turbine-resident | 0.3.4a1 | 3 | keep-up | 421,064 | 417,723 | 338,091 | 2.96 | 280,895 | 5,650.2 | - | - | 2026-11-10 |
| turbine-resident · drain | 0.3.5a1 | 3 | drain | — | 858,879 | 299,018 | 3.34 | 444,738 | - | -6.2% ⚠ | -6.0% ⚠ | 2026-11-10 |
| turbine-resident · drain | 0.3.4a1 | 3 | drain | — | 914,056 | 318,755 | 3.14 | 468,017 | - | - | - | 2026-11-10 |
How to read this table
The Turbine lines measured on several releases, newest first: a regression check, not part of the cross-product comparison above (which shows the pinned release only). Δ compares each release with the one below it: efficiency = rows/s per core (the inverse of µs CPU per record), capacity = consumed/s, only when both runs were fed beyond capacity (saturated or drain — at a keep-up point consumed/s is just the offered rate). ⚠ = worse by more than 5%, about the single-run noise: confirm with --repeat 3. Two releases measured in two sessions also differ by the host-to-host spread — same flavor, another physical host: −0.7 % consumed / −3.5 % rows/s per core in the c3 control of 2026-10-07 —, which leaves little headroom under the 5 %: re-measure both releases in ONE session (suite.py run … --version X) before acting on a ⚠. No delta is computed across a change of bench version (bench, from bench.yaml), host or run config; ⚠ deprecated rows were measured on an older bench definition — re-run that release with suite.py run turbine WORKLOAD --version X.
Host: ovh:c3-32 (ovh:c3-32@DE1, ovh:c3-32@SBG5, ovh:c3-32@UK1) · AMD EPYC-Milan Processor (2400 MHz) · 16 cores / 16 threads · 31345 MB · kernel 6.12.100+deb13-cloud-amd64.
Redpanda: rpk version: v26.2.2. Redpanda, gendata and the product under test share this host.