Collaboration performance test report

Test environment, metrics, and results for real-time concurrent editing with the Univer collaboration engine.

We tested multiple editors in one document and simultaneous edits across many documents. After warming up the documents, we measured acknowledgment latency, throughput, the percentage of planned submissions that succeeded, and the collaboration process's CPU and memory usage.

Test environment

The collaboration service and PostgreSQL ran on separate servers. The machine sizes in the results refer to the collaboration node in the diagram.

100%

PostgreSQL stores document records, changesets, snapshot records, and block references. OSS stores snapshot cell-data blocks.

All measurements are for a single collaboration node. When concurrent documents exceed its capacity, additional nodes can handle edits to different documents. They share a database and use pub/sub to synchronize rooms, presence, and edit broadcasts. See Multiple collaboration nodes.

ComponentConfiguration
Univer Collaboration SDK version1.5.2
OS and runtimeUbuntu 26.04.1 LTS, Node.js 24.14.0
StoragePostgreSQL 18.6 and Alibaba Cloud OSS
Database adapterPostgreSQL + S3 Adapter
Collaboration nodeAlibaba Cloud ECS ecs.g9i.xlarge (4 vCPUs, 16 GiB) or ecs.g9i.2xlarge (8 vCPUs, 32 GiB), Intel Xeon 6982P-C, 40 GiB ESSD
Database serverAlibaba Cloud ECS ecs.c9i.large, 2 vCPUs, 4 GiB, Intel Xeon 6982P-C, 60 GiB ESSD
Load generatorAlibaba Cloud ECS ecs.u2a-c1m2.xlarge, 4 vCPUs, 8 GiB, AMD EPYC
NetworkAlibaba Cloud Shenzhen region. Clients, collaboration node, and database communicate over the regional private network. OSS uses an internal endpoint.

Collaboration service configuration

On both node sizes, we raised runtime.maxLoadedUnits from 1,200 to 4,000 so the node could keep up to 4,000 documents in memory. Other SDK runtime settings used their defaults, including automatic worker selection. The default snapshot policy triggers after 50 revisions since the last successful snapshot or when a single changeset reaches 5 MiB.

The benchmark service uses an initialized PostgreSQL pool and OSS client. The transport attaches to the Node.js HTTP server.

TypeScript
const dbAdapter = await PostgresS3DatabaseAdapter.create({  pool,  s3: { client, bucket, prefix },});const service = new UniverCollabService({  dbAdapter,  runtime: { maxLoadedUnits: 4000 },});const endpoint = new UniverCollabEndpoint(service);const transport = createNodeTransport();transport.register(endpoint);transport.attach(server);

Test snapshots

Each document starts from a snapshot. Cell counts refer to cells containing data or formulas before warm-up.

SnapshotPopulated cellsContents
blank0Empty worksheet
10k-cellsAbout 10,000Numbers, text, and formulas
100k-cellsAbout 100,000Numbers, text, and formulas

Both populated sheets have 26 columns: 14 numeric, nine text, and three containing formulas that reference cells in the same row.

Method

Simulated clients

Each editor in the results is an independent session. Each session joins one document, submits a cell edit over HTTP, and receives the acknowledgment over WebSocket. The scripts call the collaboration API directly. Both the HTTP request and the wait for acknowledgment have a 10-second timeout.

Warm-up and sampling

Each run creates documents from the selected snapshot. Warm-up starts once all sessions have joined. The first edit loads the document on the server. A single document receives one warm-up edit. In multi-document runs, successive documents receive one to 49 edits, restarting at one for the 50th document. This staggers the starting revisions and snapshot writes.

Sampling starts after all documents have warmed up. We record first-submission latency separately during warm-up. Submission counts, latency percentiles, CPU, and memory in the tables come from the sampling period.

Edit workload

For the single-document test we increased the number of editors on blank. For each of the three snapshots we also increased the number of active documents. Each submission contains one sheet.mutation.set-range-values operation that writes a number to one cell.

ConditionOne documentMultiple documents
Editors per documentVaries by run (see results)Two in every run
Edit operationStaggered writes to each editor's own cellStaggered writes to each editor's own cell
Frequency per editorOnce every 6 secondsOnce every 2 seconds
Sampling period per run300 seconds300 seconds

Metrics

MetricDefinition
Acknowledgment latencyTime from sending an edit until the same client receives the server acknowledgment. P50, P95, and P99 are the times within which 50%, 95%, and 99% of successful submissions were acknowledged. Max is the longest successful acknowledgment time.
ThroughputSuccessful submissions divided by sampling time, in submissions per second.
Completion rateSuccessful submissions as a percentage of all planned submissions, including failed and unsent ones.
CPU mean / peak (%)Mean and highest sampled CPU usage of the collaboration process. One fully utilized core counts as 100%, so totals can exceed 100%.
Peak memory (MiB)Highest sampled resident set size (RSS) of the collaboration process.

Pass criteria

A run passes when acknowledgment latency is P99 < 1,000 ms, the completion rate is > 90%, and the collaboration process stays up until every test document passes the consistency check. Latency and completion rate are evaluated before rounding.

A load level is evaluated across the runs reported here. A single run must pass, and if there are multiple runs, every run must pass. Interrupted samples and missing metrics are marked as missing with the reason.

Interpreting the results

Node.js scripts simulate editing sessions and send cell changes directly to the API. Acknowledgment latency measures the round trip from the script to the server. It excludes browser interaction and spreadsheet rendering.

Use these measurements for performance assessment and single-node capacity planning. Performance depends on deployment, network conditions, document size, and edit type and frequency.

The highest passing load is the largest tested level that met the criteria. Mean CPU usage approached node capacity at some passing levels. Treat these levels as capacity-planning references, not sustained production targets. Leave headroom for concurrency spikes and resource fluctuations, and validate sustained capacity with representative workloads in the target environment.

Results

Summary

The tables show the highest passing load per node size and document size. Load counts editors for one document and active documents for multi-document runs. Each row gives throughput, latency, CPU, and memory at that level. See detailed results for completion rates and higher loads.

Multiple editors in one document

Each node size passed one run with 500 editors.

SnapshotNode sizeHighest passing editor countThroughput (edits/s)P95 (ms)P99 (ms)CPU mean / peak (%)Peak memory (MiB)
blank4 cores / 16 GiB50083.313.317.427.8 / 54.8777
blank8 cores / 32 GiB50083.313.918.025.6 / 100.51,322

Concurrent editing across documents

SnapshotNode sizeHighest passing document countThroughput (edits/s)P95 (ms)P99 (ms)CPU mean / peak (%)Peak memory (MiB)
blank4 cores / 16 GiB2,1002,100.029.9194.6118.9 / 332.65,829
blank8 cores / 32 GiB2,9002,900.017.856.2145.3 / 363.97,404
10k-cells4 cores / 16 GiB1,8001,800.0106.2311.9241.3 / 379.36,304
10k-cells8 cores / 32 GiB2,4002,400.026.9482.0286.3 / 749.38,071
100k-cells4 cores / 16 GiB380380.091.1200.9389.6 / 401.812,440
100k-cells8 cores / 32 GiB750750.033.572.3649.7 / 782.129,276

Load estimates for 4 cores / 8 GiB and 8 cores / 16 GiB

The 4-core/8-GiB and 8-core/16-GiB figures are estimates based on measurements from 4-core/16-GiB and 8-core/32-GiB nodes respectively. We allowed 6.5 GiB and 13.7 GiB for the collaboration process. For each workload and snapshot we selected the highest passing level in the detailed results whose peak RSS fit the budget, using the highest RSS across runs.

Multiple editors in one document
SnapshotTarget node sizeEstimated editorsPeak memory (MiB)
blank4 cores / 8 GiB500777
blank8 cores / 16 GiB5001,322
Concurrent editing across documents
SnapshotTarget node sizeEstimated active documentsPeak memory (MiB)
blank4 cores / 8 GiB2,1005,829
blank8 cores / 16 GiB2,9007,404
10k-cells4 cores / 8 GiB1,8006,304
10k-cells8 cores / 16 GiB2,4008,071
100k-cells4 cores / 8 GiB3505,413
100k-cells8 cores / 16 GiB6008,943

Sizing node pools

In a multi-node deployment, the gateway can route documents by size to separate pools. These single-node estimates assume two editors per document, each editing once every two seconds.

  • For small and medium documents, evaluate 4 cores and 8 GiB first. In the 4-core/16-GiB tests, the highest passing loads within the smaller memory budget were 2,100 blank sheets or 1,800 sheets with about 10,000 populated cells each.
  • For large documents, evaluate 8 cores with 16 or 32 GiB. With about 100,000 populated cells per document, the 8-core/32-GiB node passed at 750 documents. The largest passing load within the 16-GiB configuration's memory budget was 600.

Detailed results

Single-run rows show measured values. When combining runs at the same level, we report arithmetic means for throughput, P50, P95, P99, Max, completion rate, and mean CPU. Peak CPU and memory are the highest observed across runs. Notes give the run count, the range of per-run maxima, and the P99 range where relevant. Each run is evaluated separately.

Multiple editors in one document

Snapshot · blank
4 cores / 16 GiB
EditorsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
15025.02.94.35.412.6100.00%5.7 / 27.9690—
40066.75.210.112.429.8100.00%19.8 / 41.9781—
50083.36.613.317.440.1100.00%27.8 / 54.8777—
8 cores / 32 GiB
EditorsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
15025.03.44.45.510.7100.00%5.1 / 66.51,123—
40066.76.210.813.428.2100.00%18.2 / 75.91,196—
50083.37.513.918.036.9100.00%25.6 / 100.51,322—

Concurrent editing across documents

Snapshot · blank
4 cores / 16 GiB
Active documentsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
500500.01.11.63.725.0100.00%33.8 / 89.32,057—
1,0001,000.01.12.410.757.0100.00%61.4 / 193.13,133—
1,5001,500.01.34.840.1361.6100.00%86.3 / 279.04,134—
2,1002,100.01.729.9194.61,156.2100.00%118.9 / 332.65,8293 runs, P99 range 141.9–254.0 ms, Max range 508.2–1,543.0 ms
2,2002,200.01.870.1270.41,227.5100.00%123.5 / 369.96,050Run 1
2,2002,199.82.0274.0860.42,787.299.99%127.1 / 392.66,172Run 2
2,2002,198.82.2376.51,062.43,747.899.95%126.1 / 347.26,071Run 3, latency missed target
2,3002,298.13.2878.61,675.84,313.999.92%135.8 / 370.86,336P99 missed target
2,4002,204.82,192.52,457.52,888.64,611.591.87%163.6 / 394.87,023P99 missed target
8 cores / 32 GiB
Active documentsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
500500.01.01.32.226.1100.00%30.6 / 78.12,823—
1,0001,000.01.11.73.958.9100.00%53.8 / 163.43,961—
1,5001,500.01.22.69.651.8100.00%76.3 / 165.14,694—
2,2002,200.01.46.515.0189.2100.00%112.2 / 428.16,030—
2,6002,600.01.813.038.8334.1100.00%132.2 / 477.56,891—
2,7002,700.01.811.735.0270.5100.00%135.1 / 661.77,0263 runs, peak CPU 518.6%–661.7%, Max range 234.9–313.9 ms
2,8002,800.02.014.652.1352.0100.00%140.5 / 537.47,3313 runs, peak CPU 464.9%–537.4%, Max range 273.3–430.0 ms
2,9002,900.02.217.856.2395.2100.00%145.3 / 363.97,4041 run
3,0002,375.52,537.12,715.42,868.13,821.879.18%194.3 / 570.98,2572 runs, P99 and completion rate missed targets, Max range 3,311.4–4,332.2 ms
Snapshot · 10k-cells
4 cores / 16 GiB
Active documentsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
500500.01.13.27.7206.5100.00%71.1 / 165.82,514—
1,0001,000.01.35.116.4115.6100.00%133.6 / 330.53,879—
1,6001,600.02.042.7256.71,165.7100.00%215.2 / 371.15,677—
1,8001,800.02.7106.2311.91,197.3100.00%241.3 / 379.36,3043 runs, Max range 773.7–1,455.4 ms
1,9001,899.95.4330.2797.13,210.599.99%257.9 / 399.36,626Run 1
1,9001,891.44.6552.61,541.56,200.099.54%256.7 / 397.26,686Run 2, latency missed target
2,0001,868.42,084.52,401.92,964.07,658.393.42%284.8 / 397.67,104P99 missed target
2,1001,904.02,202.32,631.73,022.04,550.790.67%291.2 / 393.07,580P99 missed target
8 cores / 32 GiB
Active documentsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
500500.01.02.04.127.3100.00%69.3 / 139.93,408—
1,0001,000.01.12.86.830.3100.00%127.2 / 238.84,601—
2,0002,000.01.58.221.2184.4100.00%249.9 / 634.87,222—
2,2002,200.01.711.633.8244.0100.00%268.6 / 713.07,716—
2,3002,300.01.914.4197.21,043.4100.00%284.4 / 696.98,0713 runs, P99 range 39.6–368.8 ms, Max range 326.1–1,700.6 ms
2,4002,400.02.126.9482.02,287.6100.00%286.3 / 749.38,0711 run
2,5002,499.83.3600.41,228.13,346.499.99%302.5 / 744.78,308P99 missed target
2,6002,367.61,053.52,489.82,646.24,510.691.06%324.3 / 741.48,865P99 missed target
2,7002,222.72,467.72,583.52,687.63,627.482.32%334.4 / 776.59,2382 runs, P99 and completion rate missed targets, Max range 3,547.4–3,707.3 ms
Snapshot · 100k-cells
4 cores / 16 GiB
Active documentsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
100100.01.46.622.868.0100.00%91.6 / 202.22,662—
250250.01.923.056.8637.3100.00%246.3 / 396.64,046—
350350.05.176.2170.5479.8100.00%365.5 / 397.95,413—
380380.05.791.1200.9725.9100.00%389.6 / 401.812,4403 runs, Max range 615.2–856.9 ms
390390.05.8106.0275.31,339.3100.00%391.1 / 399.613,639Out of memory after sampling
400394.76.2108.1210.3675.898.67%377.4 / 398.314,226Out of memory, 1,385 failed submissions
8 cores / 32 GiB
Active documentsThroughput (edits/s)P50 (ms)P95 (ms)P99 (ms)Max (ms)Completion rateCPU mean / peak (%)Peak memory (MiB)Notes
100100.01.32.821.736.7100.00%91.8 / 200.13,392—
250250.01.25.823.0284.2100.00%242.0 / 471.15,013—
500500.01.516.233.9121.6100.00%482.6 / 716.47,721—
600600.02.026.765.6302.1100.00%575.9 / 772.18,943—
700700.02.429.964.5288.1100.00%641.4 / 782.525,692—
750750.02.633.572.3280.7100.00%649.7 / 782.129,2763 runs, Max range 266.2–295.1 ms
775725.22.732.771.0314.293.57%607.2 / 789.129,360Out of memory, 4,474 failed submissions

We continue to optimize the collaboration service to reduce CPU and memory use and improve editing performance for large documents.

How is this guide?

© 2026 DreamNum Co., Ltd.