Spark Optimization

Field-tested techniques for optimizing open-source Apache Spark 3.5.0 (and open-source Delta Lake) workloads. Every technique lists what it does, why it matters, patterns (when it helps), anti-patterns (when it does not help or hurts), the exact config/code, and impact evidence from real reproductions.

Applies to plain OSS Spark 3.5.0. Vendor-only features (Databricks Photon / predictive optimization / Delta disk cache, AWS Glue/EMR runtime tricks) are excluded or explicitly flagged. (Note: Delta Lake liquid clustering, OPTIMIZE/ZORDER, deletion vectors, CDF are open-source — GA in Delta 3.2.0 — see references/08-delta-lake.md.) Some examples run on the Oracle AI Data Platform (AIDP); AIDP-specific execution rules are in references/aidp-notes.md.

In this plugin (oracle-ai-data-platform-workbench-engineer-agent): measure/apply on AIDP via the bundled scripts/aidp_sql.py (one SparkSession per cluster -- set/revert spark.conf explicitly; see references/aidp-notes.md); pull Spark-UI stage/task metrics via the aidp-spark-debugging skill; right-size compute + set cluster-create-only configs (e.g. spark.memory.fraction) via aidp-cluster-ops; run Delta OPTIMIZE/VACUUM/ZORDER DDL via aidp-sql-ddl / aidp-table-management. Adapted from the ai-data-engineer-agent spark-optimization skill (Oracle).

Core principle

Optimization is a measurement loop, not a guess. Detect → diagnose → propose → apply → re-run on the same data → compare the same metrics.

Most Spark time is lost to moving data (shuffle), waiting on a straggler (skew), spilling (under-memory), doing redundant work (re-reads, re-materialization, repeated actions), or per-row execution overhead (wide codegen). Almost every win below reduces one of those.

When to use

• A job/notebook is slow, missing an SLA, or its cost is too high.
• Spark UI shows: a single slow stage, task p100/p50 > 2x skew, large shuffle read/write, memory/disk spill, high GC, or thousands of tiny tasks/files.
• You see code smells: union of the same table, a driver loop calling collect()/head()/count() repeatedly, MERGE on a huge table, wide groupBy (100+ fields), reading 100k+ small files, an un-filtered join with a big dimension.
• You read from / write to an Oracle database (ADW/ATP/Exadata): a single-task JDBC read, a slow or redo-saturating multi-TB write, or unsupported array/map/struct columns. (references/09-oracle-database.md)
• Before launching a large ingest/transform so the cluster + configs are right the first time.

When NOT to use: correctness bugs (use debugging), or sub-second jobs where tuning effort exceeds the payoff. Structured Streaming engine tuning (trigger interval, maxOffsetsPerTrigger, state store) is out of scope here — but each micro-batch is batch-shaped, so the join/skew/partition/file/codegen techniques below apply per batch.

How to find the opportunity → which reference to read

digraph route {
  rankdir=LR; node [shape=box];
  sym [shape=diamond, label="Dominant symptom?"];
  sym -> "references/diagnosis.md" [label="not sure / need to measure"];
  sym -> "references/02-joins.md" [label="slow join, skew, SortMergeJoin"];
  sym -> "references/04-memory-and-spill.md" [label="spill, OOM, lost executors, GC"];
  sym -> "references/cluster-sizing.md" [label="right-size compute: workers/OCPU/RAM"];
  sym -> "references/03-file-layout-io.md" [label="many small files, slow scan, listing"];
  sym -> "references/01-partitioning.md" [label="too many/few partitions, weak parallelism"];
  sym -> "references/05-codegen.md" [label="CPU-bound wide aggregation"];
  sym -> "references/06-caching-materialization.md" [label="re-reads, union, driver loop, MERGE"];
  sym -> "references/07-aqe.md" [label="want runtime adaptivity"];
  sym -> "references/08-delta-lake.md" [label="Delta table: small files, OPTIMIZE/ZORDER/clustering, VACUUM, deletion vectors, CDF"];
  sym -> "references/09-oracle-database.md" [label="read/write Oracle DB / ADW / Exadata (JDBC parallelism, redo, complex types)"];
}

Then consult references/config-matrix.md for the exact key, default, impact, and where it can be set (notebook / cluster-create-only / non-modifiable), and references/quick-reference.md for the impact-ranked checklist.

Reference index

The 80/20: highest-impact moves first

Ranked by typical payoff (full details + caveats in the referenced files):

1. Eliminate avoidable shuffle on joins — raise spark.sql.autoBroadcastJoinThreshold so a small/medium side broadcasts; or semi-join pre-filter the big dimension down to broadcastable size. (02-joins.md)
2. Fix the small-file problem — compact sources / size output files; tune maxPartitionBytes + openCostInBytes. Listing+opening 100k+ files can cost more than the data. (03-file-layout-io.md)
3. Right-size memory to stop spill — raise spark.memory.fraction (cluster-create-only on AIDP) when there are no user data structures; spill to disk is a silent multiplier. (04-memory-and-spill.md, cluster-sizing.md)
4. Address skew — AQE skew-join, or semi-join pre-filter + broadcast; a p100/p50 > 2x join stage serializes the whole stage on a straggler. (02-joins.md, 07-aqe.md)
5. Avoid redundant work — don't union the same table, don't loop collect()/head() on the driver, treat intermediates as views (don't materialize), prefer JOIN over MERGE for full reloads. (06-caching-materialization.md)
6. Compress shuffle/spill/output with zstd — spark.io.compression.codec=zstd + spark.sql.parquet.compression.codec=zstd: tight like gzip, fast like snappy; cuts spill-disk and output size. (03-file-layout-io.md)
7. Tune parallelism — shuffle.partitions ≈ 2–3× total cores for large shuffles; pair with AQE coalescing to avoid small output files. (01-partitioning.md)
8. Constrain codegen on very wide aggregations — lower spark.sql.codegen.maxFields when a wide groupBy is CPU-bound; generated code can be slower than the fallback. (05-codegen.md)

Workflow

1. Measure — run the workload, collect stage metrics + task quantiles (diagnosis.md). On AIDP, see aidp-notes.md for collecting Spark UI data.
2. Diagnose — read explain(True)/the SQL plan; map the dominant symptom to a technique.
3. Propose — pick the technique; check config-matrix.md for where the config can be set.
4. Apply — change code or config. On a shared cluster, remember: one SparkSession per cluster — a spark.conf.set leaks to other notebooks; revert it explicitly (aidp-notes.md).
5. Re-run on the same cluster + data, prove the workload is logically identical (same rows/columns/result), and compare the same metrics (wall time, shuffle, spill, task skew).

Common mistakes

• Comparing wall time of two runs that aren't logically identical (different data/cache state). Prove equivalence first.
• Reading summed task metrics as wall-clock — stage duration is start→end timestamps; cumulative task time is summed across tasks.
• Counting AQE-skipped stages as time spent.
• repartition() "to be safe" — it forces a shuffle; only repartition with a reason (see 01-partitioning.md).
• Lowering AQE skew/advisory partition size so far that you create a new small-file problem.
• Setting a cluster-create-only config (e.g. spark.memory.fraction) from a notebook and assuming it took effect — it didn't. Verify with spark.conf.get / the Environment tab (config-matrix.md).