Search before asking
Description
checkpointing.interval is not a Flink configuration key, so the Python e2e tests that set it are running with checkpointing disabled.
The tests set this three-line block in 11 places across 5 files, for example python/flink_agents/e2e_tests/e2e_tests_integration/flink_intergration_test.py:53-55:
config.set_string("state.backend.type", "rocksdb")
config.set_string("checkpointing.interval", "1s")
config.set_string("restart-strategy.type", "disable")The real option is execution.checkpointing.interval, defined at CheckpointingOptions.java:512. The bare checkpointing. form is not registered anywhere, and it is not wired up as a deprecated or fallback key either, so Flink ignores it silently. execution.checkpointing.interval has .noDefaultValue() (CheckpointingOptions.java:514), so when the key is ignored no periodic checkpointing is scheduled at all.
The net effect is that these tests configure a RocksDB state backend that is never snapshotted, while reading as though they exercise checkpointing.
Verified by grepping the Flink source at every version this repo builds against (dist/ has flink-1.20 through flink-2.3). On release-1.20, release-2.2.0, release-2.3 and master, key("execution.checkpointing.interval") is present and key("checkpointing.interval") returns nothing. Also checked that PyFlink does not remap the short form on the Python side: no hits for "checkpointing.interval" anywhere under pyflink/.
Affected files:
python/flink_agents/e2e_tests/e2e_tests_integration/execute_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/flink_intergration_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/workflow_memory_remote_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/e2e_tests_mcp/mcp_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/mock_chat_model_test.py
Suggested fix: rename the key to execution.checkpointing.interval in all 11 places. Worth noting that the fix may not be purely mechanical. These tests have never actually checkpointed, so turning checkpointing on for the first time could surface real failures, and those failures would be the more interesting half of this issue.
Are you willing to submit a PR?
Search before asking
Description
checkpointing.intervalis not a Flink configuration key, so the Python e2e tests that set it are running with checkpointing disabled.The tests set this three-line block in 11 places across 5 files, for example
python/flink_agents/e2e_tests/e2e_tests_integration/flink_intergration_test.py:53-55:The real option is
execution.checkpointing.interval, defined atCheckpointingOptions.java:512. The barecheckpointing.form is not registered anywhere, and it is not wired up as a deprecated or fallback key either, so Flink ignores it silently.execution.checkpointing.intervalhas.noDefaultValue()(CheckpointingOptions.java:514), so when the key is ignored no periodic checkpointing is scheduled at all.The net effect is that these tests configure a RocksDB state backend that is never snapshotted, while reading as though they exercise checkpointing.
Verified by grepping the Flink source at every version this repo builds against (
dist/hasflink-1.20throughflink-2.3). Onrelease-1.20,release-2.2.0,release-2.3andmaster,key("execution.checkpointing.interval")is present andkey("checkpointing.interval")returns nothing. Also checked that PyFlink does not remap the short form on the Python side: no hits for"checkpointing.interval"anywhere underpyflink/.Affected files:
python/flink_agents/e2e_tests/e2e_tests_integration/execute_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/flink_intergration_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/workflow_memory_remote_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/e2e_tests_mcp/mcp_test.pypython/flink_agents/e2e_tests/e2e_tests_integration/mock_chat_model_test.pySuggested fix: rename the key to
execution.checkpointing.intervalin all 11 places. Worth noting that the fix may not be purely mechanical. These tests have never actually checkpointed, so turning checkpointing on for the first time could surface real failures, and those failures would be the more interesting half of this issue.Are you willing to submit a PR?