etcd Robustness Testing

This document describes the robustness testing framework for etcd, a distributed key-value store. The purpose of these tests is to rigorously validate that etcd maintains its KV API guarantees and watch API guarantees under a wide range of conditions and failures.

Robustness vs Antithesis tests

Antithesis runs the robustness tests inside their deterministic simulation testing environment and fault injection.

For more details on Antithesis integration, see the antithesis directory.

Robustness track record

Maintaining Bug Reproducibility During Non-Trivial Changes

When performing large non-trivial changes to the robustness testing framework, it is critical to ensure that we do not lose the ability to reproduce previously discovered bugs. The track record table above documents known correctness issues, and many include specific reproduction commands (e.g., make test-robustness-issue14370).

To prevent regressions, we must ensure that the latest version of the robustness framework remains capable of reproducing old bugs. We manually track this capability in the “Last reproduction commit” column.

Best Practices:

• Establish Baseline: Before starting a large non-trivial change, run all reproducible test cases listed in the track record table.
• Verify Reproducibility: After completing the change, verify that all previously reproducible bugs can still be detected.
• Update Tracking: Refresh the “Last reproduction commit” column with commit hash and it's creation date to confirm the new framework version works.
• Update Commands: If the change affects test execution, update the reproduction commands accordingly.
• Gate Completion: Consider the change incomplete until all regression tests continue to catch their target bugs.

This ensures that improvements to the testing framework do not inadvertently reduce our ability to detect known failure modes.

How Robustness Tests Work

Robustness tests compare the etcd cluster behavior against a simplified model of its expected behavior. These tests cover various scenarios, including:

• Different etcd cluster setups: Cluster sizes, configurations, and deployment topologies.
• Client traffic types: Variety of key-value operations (puts, ranges, transactions) and watch patterns.
• Failures: Network partitions, node crashes, disk failures, and other disruptions.

Test Procedure:

1. Cluster Creation: A new etcd cluster is created with the specified configuration.
2. Traffic and Failures: Client traffic is generated and sent to the cluster while failures are injected.
3. History Collection: All client operations and their results are recorded.
4. Validation: The collected history is validated against the etcd model and a set of validators to ensure consistency and correctness.
5. Report Generation: If a failure is detected then a detailed report is generated to help diagnose the issue. This report includes information about the client operations and etcd data directories.

Key Concepts

Distributed System Terminology

• Consensus: A process where nodes in a distributed system agree on a single data value. Etcd uses the Raft algorithm to achieve consensus.
• Strict vs Eventual consistency:
  • Strict Consistency: All components see the same data at the same time after an update.
  • Eventual Consistency: Components may temporarily see different data after an update but converge to the same view eventually.
• Consistency Models (https://jepsen.io/consistency)
  • Single-Object Consistency Models:
    • Sequential Consistency: A strong single-object model. Operations appear to take place in some total order, consistent with the order of operations on each individual process.
    • Linearizable Consistency: The strongest single-object model. Operations appear to happen instantly and in order, consistent with real-time ordering.
  • Transactional Consistency Models
    • Serializable Consistency: A transactional model guaranteeing that transactions appear to occur in some total order. Operations within a transaction are atomic and do not interleave with other transactions. It's a multi-object property, applying to the entire system, not just individual objects.
    • Strict Serializable Consistency: The strongest transactional model. Combines the total order of serializability with the real-time ordering constraints of linearizability.

Etcd provides strict serializability for KV operations and eventual consistency for Watch.

Etcd Guarantees

• Key-value API operations https://etcd.io/docs/latest/learning/api_guarantees/#kv-apis
• Watch API guarantees https://etcd.io/docs/latest/learning/api_guarantees/#watch-apis

Kubernetes Integration

• Implicit Kubernetes-ETCD Contract: Defines how Kubernetes uses etcd to store cluster state.
• ResourceVersion: A string used by Kubernetes to track resource versions, corresponding to etcd revisions.
• Sharding resource types: Kubernetes treats each resource type as a totally independent entity. It allows sharding each resource type into a separate etcd cluster.

Running locally

1. Build etcd with failpoints

   make gofail-enable
   make build
   make gofail-disable
   

2. Run the tests

   make test-robustness
   

   Optionally, you can pass environment variables:

   • GO_TEST_FLAGS - to pass additional arguments to go test. It is recommended to run tests multiple times with failfast enabled. this can be done by setting GO_TEST_FLAGS='--count=100 --failfast'.
   • EXPECT_DEBUG=true - to get logs from the cluster.
   • RESULTS_DIR - to change the location where the results report will be saved.
   • PERSIST_RESULTS - to persist the results report of the test. By default this will not be persisted in the case of a successful run.
   • TRACING_SERVER_ADDR - to export Open Telemetry traces from test runs to the collector running at given address, for example: localhost:4317

Re-evaluate existing report

Robustness test validation is constantly changing and improving. Errors in the etcd model could be causing false positives, which makes the ability to re-evaluate the reports after we fix the issue important.

Note: Robustness test report format is not stable, and it's expected that not all old reports can be re-evaluated using the newest version.

1. Identify the location of the robustness test report.

   Note: By default robustness test report is only generated for failed test.

   • For local runs: this would be by identifying log line, in the following example that would be /tmp/TestRobustnessExploratory_Etcd_HighTraffic_ClusterOfSize1:

     logger.go:146: 2024-04-08T09:45:27.734+0200 INFO    Saving robustness test report   {"path": "/tmp/TestRobustnessExploratory_Etcd_HighTraffic_ClusterOfSize1"}
     

   • For remote runs on CI: you need to go to the Prow Dashboard, go to a build, download one of the Artifacts (artifacts/results.zip), and extract it locally.

     

     

     

     Each directory will be prefixed by TestRobustness each containing a robustness test report.

     

     Pick one of the directories within the archive corresponding to the failed test scenario. The largest directory by size usually corresponds to the failed scenario. If you are not sure, you may check which scenario failed in the test logs.

2. Copy the robustness report directory into the testdata directory.

   The testdata directory can contain multiple robustness test reports. The name of the report directory doesn‘t matter, as long as it’s unique to prevent clashing with reports already present in testdata directory. For example, the path for history.html file could look like $REPO_ROOT/tests/robustness/testdata/v3.5_failure_24_April/history.html.

3. Run make test-robustness-reports to validate all reports in the testdata directory.

Analysing failure

If robustness tests fail, we want to analyse the report to confirm if the issue is on etcd side. The location of the directory with the report is mentioned in the Saving robustness test report log. Logs from report generation should look like:

    logger.go:146: 2024-05-08T10:42:54.429+0200 INFO    Saving robustness test report   {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550"}
    logger.go:146: 2024-05-08T10:42:54.429+0200 INFO    Saving member data dir  {"member": "TestRobustnessRegressionIssue14370-test-0", "path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/server-TestRobustnessRegressionIssue14370-test-0"}
    logger.go:146: 2024-05-08T10:42:54.430+0200 INFO    no watch operations for client, skip persisting {"client-id": 1}
    logger.go:146: 2024-05-08T10:42:54.430+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-1/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.430+0200 INFO    Saving watch operations {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-2/watch.json"}
    logger.go:146: 2024-05-08T10:42:54.431+0200 INFO    no KV operations for client, skip persisting    {"client-id": 2}
    logger.go:146: 2024-05-08T10:42:54.431+0200 INFO    no watch operations for client, skip persisting {"client-id": 3}
    logger.go:146: 2024-05-08T10:42:54.431+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-3/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.433+0200 INFO    no watch operations for client, skip persisting {"client-id": 4}
    logger.go:146: 2024-05-08T10:42:54.433+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-4/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.434+0200 INFO    no watch operations for client, skip persisting {"client-id": 5}
    logger.go:146: 2024-05-08T10:42:54.434+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-5/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.435+0200 INFO    no watch operations for client, skip persisting {"client-id": 6}
    logger.go:146: 2024-05-08T10:42:54.435+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-6/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.437+0200 INFO    no watch operations for client, skip persisting {"client-id": 7}
    logger.go:146: 2024-05-08T10:42:54.437+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-7/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.438+0200 INFO    no watch operations for client, skip persisting {"client-id": 8}
    logger.go:146: 2024-05-08T10:42:54.438+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-8/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.439+0200 INFO    no watch operations for client, skip persisting {"client-id": 9}
    logger.go:146: 2024-05-08T10:42:54.439+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-9/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.440+0200 INFO    no watch operations for client, skip persisting {"client-id": 10}
    logger.go:146: 2024-05-08T10:42:54.440+0200 INFO    Saving operation history        {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/client-10/operations.json"}
    logger.go:146: 2024-05-08T10:42:54.441+0200 INFO    Saving visualization    {"path": "/tmp/TestRobustnessRegression_Issue14370/1715157774429416550/history.html"}

The report follows the hierarchy:

• server-* - etcd server data directories, can be used to verify disk/memory corruption.
  • member
    • wal - Write Ahead Log (WAL) directory, that can be analysed using etcd-dump-logs command line tool available in tools directory.
    • snap - Snapshot directory, includes the bbolt database file db, that can be analysed using etcd-dump-db command line tool available in tools directory.
• client-* - Client request and response dumps in json format.
  • watch.json - Watch requests and responses, can be used to validate watch API guarantees.
  • operations.json - KV operation history
• history.html - Visualization of KV operation history, can be used to validate KV API guarantees.

Example analysis of a linearization issue

Let's reproduce and analyse robustness test report for issue #14370. To reproduce the issue by yourself run make test-robustness-issue14370. After a couple of tries robustness tests should fail with a log Linearization illegal and save the report locally.

Example:

    logger.go:146: 2025-08-01T22:54:26.550+0900 INFO Validating linearizable operations {"timeout": "5m0s"}
    logger.go:146: 2025-08-01T22:54:26.755+0900 ERROR Linearization illegal {"duration": "205.05225ms"}
    logger.go:146: 2025-08-01T22:54:26.755+0900 INFO Skipping other validations as linearization failed
    main_test.go:122: linearization: illegal
    logger.go:146: 2025-08-01T22:54:26.756+0900 INFO Saving robustness test report {"path": "/tmp/TestRobustnessRegression_Issue14370/1754056466755991000"}
    ...
    logger.go:146: 2025-08-01T22:54:26.850+0900 INFO Saving visualization {"path": "/tmp/TestRobustnessRegression_Issue14370/1754056466755991000/history.html"}
    logger.go:146: 2025-08-01T22:54:26.878+0900 INFO killing server... {"name": "TestRobustnessRegressionIssue14370-test-0"}
    logger.go:146: 2025-08-01T22:54:26.878+0900 INFO stopping server... {"name": "TestRobustnessRegressionIssue14370-test-0"}
    logger.go:146: 2025-08-01T22:54:26.886+0900 INFO stopped server. {"name": "TestRobustnessRegressionIssue14370-test-0"}

Linearization issues are easiest to analyse via history visualization. Open /tmp/TestRobustnessRegression_Issue14370/1754056466755991000/history.html file in your browser. Jump to the error in linearization by clicking [ jump to first error ] on the top of the page.

You should see a graph similar to the one on the image below.

The last correct request (connected with the grey line) is a Put request that succeeded and got revision 168. All following requests are invalid (connected with red line) as they have revision 167. Etcd guarantees that revision is non-decreasing, so this shows a bug in etcd as there is no way revision should decrease. This is consistent with the root cause of #14370 as it was an issue with the process crash causing the last write to be lost.

Example analysis of a watch issue

Let's reproduce and analyse robustness test report for issue #15271. To reproduce the issue by yourself run make test-robustness-issue15271. After a couple of tries robustness tests should fail with a logs Broke watch guarantee and save report locally.

Example:

    logger.go:146: 2024-05-08T10:50:11.301+0200 INFO    Validating linearizable operations      {"timeout": "5m0s"}
    logger.go:146: 2024-05-08T10:50:15.754+0200 INFO    Linearization success   {"duration": "4.453346487s"}
    logger.go:146: 2024-05-08T10:50:15.754+0200 INFO    Validating watch
    logger.go:146: 2024-05-08T10:50:15.849+0200 ERROR   Broke watch guarantee   {"guarantee": "ordered", "client": 4, "revision": 3}
    validate.go:45: Failed validating watch history, err: broke Ordered - events are ordered by revision; an event will never appear on a watch if it precedes an event in time that has already been posted
    logger.go:146: 2024-05-08T10:50:15.849+0200 INFO    Validating serializable operations
    logger.go:146: 2024-05-08T10:50:15.866+0200 INFO    Saving robustness test report   {"path": "/tmp/TestRobustnessRegression_Issue15271/1715158215866033806"}

Watch issues are easiest to analyse by reading the recorded watch history.

Watch history is recorded for each client separated in different subdirectory under /tmp/TestRobustnessRegression_Issue15271/1715158215866033806.

Open watch.json for the client mentioned in the log Broke watch guarantee. For client 4 that broke the watch guarantee open /tmp/TestRobustnessRegression_Issue15271/1715158215866033806/client-4/watch.json.

Each line consists of json blob corresponding to a single watch request sent by the client. Look for events with Revision equal to revision mentioned in the first log with Broke watch guarantee, in this case, look for "Revision":3,. You should see watch responses where the Revision decreases like ones below:

{"Events":[{"Type":"put-operation","Key":"key5","Value":{"Value":"793","Hash":0},"Revision":799,"IsCreate":false,"PrevValue":null}],"IsProgressNotify":false,"Revision":799,"Time":3202907249,"Error":""}
{"Events":[{"Type":"put-operation","Key":"key4","Value":{"Value":"1","Hash":0},"Revision":3,"IsCreate":true,"PrevValue":null}, ...

Up to the first response, the Revision of events only increased up to a value of 799. However, the following line includes an event with Revision equal 3. If you follow the revision throughout the file you should notice that watch replayed revisions for a second time. This is incorrect and breaks Ordered watch API guarantees. This is consistent with the root cause of #15271 where the member reconnecting to cluster will resend revisions.