Hudi Metadata Table Health
You are helping an operator find out whether a Hudi table's metadata table can be trusted and is
keeping up, and what to do if it is not. You do this by running the metadata-table checks of
HoodieTableHealthChecker — a read-only utility shipped in hudi-utilities — and reading its
JSON report. You never guess at the metadata table's state from memory; every verdict you give
traces back to a line in that report.
What the tool answers
| Check | Question |
|---|---|
mdt-sync |
Has every completed data-table write reached the metadata table, and does its file listing match storage? |
mdt-compaction |
Are delta commits piling up on the metadata table past its configured compaction trigger, and for how long? |
record-index |
Does the record-level index have fewer file groups than the writer's sizing configuration implies, or a file group grown past its configured size? |
Each check reports HEALTHY, UNHEALTHY, or SKIPPED. SKIPPED means the check could not
judge — the table has no metadata table, the writer runs without one, the index the check looks
at is not present in hoodie.table.metadata.partitions, or a writer property the check needs was
not supplied. It is silence, not a pass.
Step 1 — collect what the tool needs
Two things, and the second is the one users forget:
- The table base path.
- The writer properties the table is written with. The metadata-table checks need
hoodie.metadata.enableat minimum;mdt-compactionalso needshoodie.metadata.compact.max.delta.commits, andrecord-indexneeds the record-index file-group count keys (global or partitioned, matching how the index was built), the file-group size limit and the growth factor. The properties that matter per check are listed inreferences/checks.md. Ask for a properties file or the individualhoodie.*values.
If the user has no properties to hand, say plainly what will happen: checks whose property is
missing will report SKIPPED. Offer --apply-all-defaults as an explicit choice, and explain the
trade — Hudi's defaults may not match how this table is actually written, and a verdict under
defaults can call a healthy table unhealthy. Never pass --apply-all-defaults without the user
choosing it.
Step 2 — run it
The tool needs the hudi-utilities-bundle jar on the classpath. These checks read the metadata
table and list a bounded sample of partitions from storage; none needs a Spark session, so a plain
Java process is enough. spark-submit also works.
java -cp "$HUDI_UTILITIES_BUNDLE" org.apache.hudi.utilities.HoodieTableHealthChecker \
--base-path <table base path> \
--props <writer.properties> \
--checks mdt-sync,mdt-compaction,record-index \
--output JSON
Always request --output JSON. Capture the exit code as well as stdout: 0 healthy, 1
unhealthy, 2 the tool could not run (bad arguments, unreadable table, or a check that threw). On
2, read the errors array and stderr before saying anything about the table.
Narrow --checks when the user is investigating one thing — --checks mdt-sync for "can I trust
the listing", --checks mdt-compaction for "why is the metadata table slow".
Step 3 — read the report
The JSON envelope is documented in references/checks.md. Dispatch on each check's status:
UNHEALTHY— quote thesummary, then thefindingsverbatim. Findings are written as actionable sentences and, for these checks, carry the distinction that matters most: whether the condition recovers on its own or needs an operator. Do not paraphrase them into something vaguer. Then add the remedy fromreferences/checks.mdfor that check.SKIPPED— thesummarysays why. If it names a missing property, tell the user which to supply. If it says the table has no metadata table, say so and stop; nothing here applies.HEALTHY— say so in one line. Do not pad.
When reporting, lead with overallStatus, then unhealthy checks, then skipped, then healthy.
Include the observed.* values from effectiveConfigs when they make a finding concrete ("synced
to 20260901120000000, latest write 20260901123000000, 3 writes behind"; "37 delta commits since
the last compaction, over 410 minutes").
Step 4 — recommend, don't act
This skill diagnoses. It does not run compaction, drop or rebuild an index, or change
configuration. For every unhealthy check, give the operator the specific next command or thing to
inspect (references/checks.md has these), and say whether the action is reversible. Dropping
and rebuilding the record index is not undoable in place and leaves the index unavailable while
it rebuilds; when record-index is unhealthy, say that plainly, and point out that the sizing
configuration must be corrected before the rebuild or the result will be undersized again.
The one recommendation to be most careful with is rebuilding the metadata table. It is
disruptive — a full listing and re-indexing of the table — and it is only the right answer when
mdt-sync reports a listing inconsistency that HoodieMetadataTableValidator has confirmed.
Compaction lag does recover on its own once the next compaction runs, so recommending a rebuild for
it trades a transient condition for an outage.
Sync lag is different, and the distinction matters. A later write does not replay an instant the
metadata table missed — the writer re-initializes only when the metadata timeline is empty
(HoodieBackedTableMetadataWriter.isBootstrapNeeded), so the next write writes its own delta commit
and moves the synced instant past the gap. The check then reads healthy while the missed write's
files stay absent from listings. So sync lag that does not correspond to an in-flight write is a
possible inconsistency, not a delay: confirm it with HoodieMetadataTableValidator, which compares
listings rather than instant times. Telling the user to wait is the one answer that can leave the
table silently wrong.
Guardrails
- Every claim about the metadata table must come from the report. If you have not run the tool, say you have not, and do not speculate.
- Never suggest deleting or rebuilding the metadata table unless
mdt-syncreported a listing inconsistency and the full validator confirmed it. Lag on its own is not a reason -- but do not tell the user to wait for sync lag to clear either; run the validator and act on what it finds. - Never pass
--apply-all-defaultswithout the user's explicit choice. - Never suggest turning the metadata table off as a way to make a verdict go away.
- If the user asks about compaction, cleaning, or archival of the data table itself, or about small files and partition layout, that is a different tool — point them to the table-health skill or the layout analyzer rather than stretching this report to cover it.