building-iceberg-tables

Design and operate Apache Iceberg tables — partitioning and hidden partitioning, partition/schema evolution, snapshots and time travel, compaction and small-file cleanup, and MERGE/upsert for lakehous

SKILL.md

GitHub

Building Iceberg Tables

When to use

  • Creating or maintaining Apache Iceberg tables on a lakehouse.
  • Choosing partitioning, or evolving partitioning/schema without rewrites.
  • Managing snapshots, time travel, compaction, and small files.
  • Do NOT use for Delta-specific work (use engineering-databricks-pipelines).

Workflow

- [ ] Partition by query filter columns; use hidden partition transforms
- [ ] Use MERGE for idempotent upserts
- [ ] Schedule compaction (rewrite_data_files) to fix small files
- [ ] Expire old snapshots + remove orphan files to control metadata/storage
- [ ] Evolve partitioning/schema by field ID (no data rewrite)
  1. Partition on filter columns using hidden partition transforms (days(ts), bucket(N, id)), so queries prune without users adding derived partition columns.
  2. Idempotent writes via MERGE INTO keyed on the business key.
  3. Compact regularly — streaming/small-batch writes create many small files; rewrite_data_files restores read performance.
  4. Maintain metadata — expire old snapshots and remove orphan files, or snapshot history and storage grow without bound.
  5. Evolve freely — Iceberg tracks columns/partitions by ID, so add/drop/rename and even partition-spec changes need no data rewrite.

Patterns

Create with hidden partitioning + MERGE upsert:

CREATE TABLE lake.db.orders (order_id BIGINT, customer_id BIGINT, amount DECIMAL, ordered_at TIMESTAMP)
USING iceberg PARTITIONED BY (days(ordered_at));

MERGE INTO lake.db.orders t USING staging s ON t.order_id = s.order_id
WHEN MATCHED THEN UPDATE SET *
WHEN NOT MATCHED THEN INSERT *;

Maintenance (Spark procedures):

CALL lake.system.rewrite_data_files('db.orders');       -- compact small files
CALL lake.system.expire_snapshots('db.orders', TIMESTAMP '2026-08-01 00:00:00');
CALL lake.system.remove_orphan_files(table => 'db.orders');

Time travel — read a prior snapshot for audit or recovery: SELECT * FROM lake.db.orders VERSION AS OF <snapshot_id>.

Partition evolution — ALTER TABLE ... ADD PARTITION FIELD bucket(16, customer_id) applies to new data only; old data stays valid.

Common pitfalls

  • No compaction on streaming tables — small-file explosion tanks read speed.
  • Never expiring snapshots — metadata and storage grow unbounded; schedule expiry within your time-travel retention.
  • Over-partitioning (e.g. by hour on low volume) — too many tiny partitions; match granularity to data size.
  • Adding explicit derived partition columns — defeats the point of hidden partitioning; use transforms.
  • Blind overwrite instead of MERGE — loses idempotency; MERGE by key.
  • Removing orphan files with a too-short window — can delete files in-flight writers still need; use a safe cutoff.