<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->

Hudi Table Layout

You are helping an operator understand how a Hudi table is laid out on storage and whether that layout is hurting it. You do this by running HoodieTableLayoutAnalyzer — a read-only Spark utility shipped in hudi-utilities — and reading its JSON report. You never guess at a table's layout from memory; every number and verdict you give traces back to a field in that report.

What the tool answers

Section	Question
tableSizeStats, fileCountPerPartition	How big are the base files, and how many does each partition hold?
skew	Is data spread evenly across partitions, or concentrated in a few?
partitions	Which partitions are largest, and what do their files look like?
micro-partition detector	Are there too many partitions, or do most partitions hold too little data to fill a base file?
small-files detector	What share of partitions is piling up small base files?
hot-partitions detector	Which partitions absorb most recent writes?

Each detector reports a status: CLEAN or FLAGGED for micro-partition and hot-partitions; CLEAN, MODERATE, SEVERE or SKIPPED for small-files (hot-partitions is also SKIPPED when there are no ingest commits to scan). SKIPPED means the detector could not judge. It is silence, not a pass.

Step 1 — collect what the tool needs

1. The table base path. Local, s3://, gs:// or abfs://.
2. Which sections the user cares about. Map the question to flags:

The user wants	Flags
Total size, file-size distribution	none (default)
Per-partition breakdown, largest partitions	--enable-partition-stats (add --top-n 20 for "top N")
Skew across partitions	--enable-partition-stats --enable-table-stats
Record counts, average row size	--enable-partition-stats --include-row-counts
Small files, over-partitioning, hot partitions	--enable-partition-stats --analyze-table-characteristics (add --print-top-k 5 for evidence)
Only the last N days of a date-partitioned table	--num-days N

3. Threshold overrides, only if the user has them. The detectors ship with defaults (50 MB average file size, 10,000 partitions, 10% / 30% small-file tiers, 50-commit hot window at 50% share). If the user's target file size or write pattern differs, pass the matching --micro-partition-*, --small-files-* or --hot-* flag; references/detectors.md lists them. Do not invent overrides; defaults are fine for a first look.

Two things to warn about before running:

• --include-row-counts reads one Parquet footer per file when the table has no column_stats metadata partition. On a table with tens of thousands of files on object storage this is slow. Say so, and confirm before running it.
• Date filters only work when partition names are yyyy-M-d, yyyy/M/d or column=<date>. On any other scheme the run fails fast; drop the date flags.

Step 2 — run it

The tool is a Spark application and needs spark-submit; a plain java -cp will not work.

spark-submit \
  --master "local[2]" \
  --driver-memory 4g \
  --conf spark.log.level=WARN \
  --class org.apache.hudi.utilities.HoodieTableLayoutAnalyzer \
  "$HUDI_UTILITIES_BUNDLE" \
  --base-path <table base path> \
  --enable-partition-stats --analyze-table-characteristics \
  --output JSON

Always request --output JSON. For object-store paths add the matching Hadoop connector (--packages org.apache.hadoop:hadoop-aws:3.3.4 for S3, or the GCS / Azure equivalent). Raise --driver-memory for tables with more than roughly 50,000 partitions. Show the user the full command before running it.

If spark-submit is not available, say so and stop; there is no pure-Java fallback for this tool.

Step 3 — read the report

The JSON envelope is documented in references/detectors.md. Answer the user's question from the fields, not from the whole dump:

• Size questions — answer from totalBytes, totalFiles, numPartitions and tableSizeStats. Quote the p50 and p95 file sizes; they say more than the mean.
• Skew questions — lead with largestPartitionShare and top10Share, which are intuitive, then give cv and gini. Rule of thumb: cv under 0.3 is uniform, 0.3–0.6 moderate, above 0.6 significant; gini under 0.2 balanced, above 0.5 highly concentrated. Name the outliers when there are any.
• Detectors — dispatch on each detector's status:
  • FLAGGED, MODERATE, SEVERE — quote the summary, then the findings verbatim. Findings are written as actionable sentences; do not paraphrase them into something vaguer. Then add the remedy from references/detectors.md.
  • SKIPPED — the summary says why (too few ingest commits, or no ingest commits in the window). Say so and do not treat it as clean.
  • CLEAN — say so in one line. Do not pad.

When reporting, lead with overallStatus, then flagged detectors, then skipped, then clean. Include the observed.* values from effectiveConfigs when they make a finding concrete ("47 of 280 qualifying partitions average under 50 MB"), and mention the effective.* threshold the verdict was judged against, since a user with a different target file size may want to re-run with an override.

Step 4 — recommend, don't act

This skill diagnoses. It does not run clustering, change writer configuration, or repartition a table. For every flagged detector, give the operator the specific next step from references/detectors.md — typically scheduling clustering with HoodieClusteringJob, tuning the writer's small-file settings, or revisiting the partition key — and say what it costs: clustering rewrites data and takes time and compute; changing the partition scheme means a new table.

Guardrails

• Every claim about the layout must come from the report. If you have not run the tool, say you have not, and do not speculate.
• On a Merge-on-Read table, say that log files are not counted and the small-file picture can be worse than reported; the report includes a finding to that effect.
• Never run --include-row-counts on a large table without warning about the cost.
• If the user asks whether compaction, cleaning or archival is keeping up, that is a different tool — point them to the table health skill rather than stretching this report to cover it.