spark-ignis 0.2.2

ESLint for Apache Spark jobs — analyze event logs, diagnose performance issues

Ignis

ESLint for Apache Spark jobs. Point it at an event log and get actionable diagnostics for data skew, shuffle size, spill, and bad partitioning.

$ ignis analyze /path/to/spark-event-log

──────────────────── ignis  my-spark-app ────────────────────

2 issue(s) found

  Severity   Rule              Stage  Message
 ────────────────────────────────────────────────────────────
  WARNING    data-skew             2  Stage 2 ('groupBy at job.py:42'):
                                      max task 42,300ms vs median 1,800ms (23.5x ratio)
  WARNING    partition-count       3  Stage 3 ('join at job.py:71'):
                                      2 shuffle partition(s) across 8 executor core(s)
                                      — cluster is under-utilized

╭───────────────────── data-skew — Stage 2 ──────────────────╮
│ Repartition before the shuffle with a higher partition     │
│ count, or salt the join/groupBy key to spread work across  │
│ more tasks.                                                │
╰────────────────────────────────────────────────────────────╯
╭─────────────────── partition-count — Stage 3 ──────────────╮
│ Raise spark.sql.shuffle.partitions to at least 16          │
│ (2× your 8 executor cores).                                │
╰────────────────────────────────────────────────────────────╯

Installation

pip install spark-ignis            # core only
pip install "spark-ignis[s3]"      # + AWS S3
pip install "spark-ignis[gcs]"     # + Google Cloud Storage
pip install "spark-ignis[azure]"   # + Azure Data Lake Storage

Or install from source:

git clone https://github.com/skatz1990/ignis
cd ignis
python3 -m venv .venv && source .venv/bin/activate
pip install -e .               # local files only
pip install -e ".[s3]"         # + AWS S3
pip install -e ".[gcs]"        # + Google Cloud Storage
pip install -e ".[azure]"      # + Azure Data Lake Storage

Usage

# Analyze a local event log (terminal output, exits 1 if issues found)
ignis analyze /path/to/spark-event-log

# Analyze directly from cloud storage
ignis analyze s3://my-bucket/spark-logs/application_1234_0001
ignis analyze gs://my-bucket/spark-logs/application_1234_0001
ignis analyze abfs://my-container/spark-logs/application_1234_0001

# Machine-readable JSON output — pipe to jq, store in CI artifacts
ignis analyze s3://my-bucket/spark-logs/application_1234_0001 --output json

# List all rules with their thresholds
ignis rules

Exits 0 if no issues are found, 1 if any are — in both terminal and JSON modes.

Spark event logs are standard NDJSON files (Spark 3.x) or zstd-compressed directories (Spark 4.0+). Databricks writes them to DBFS, S3, GCS, or ADLS after each job.

Cloud storage

AWS S3

pip install -e ".[s3]"
ignis analyze s3://my-bucket/spark-logs/application_1234_0001

Credentials from the standard AWS chain:

Source	How
Environment variables	AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY
Named profile	AWS_PROFILE=my-profile ignis analyze s3://...
Instance role (EC2/ECS)	No configuration needed
SSO	aws sso login then run ignis normally

Google Cloud Storage

pip install -e ".[gcs]"
ignis analyze gs://my-bucket/spark-logs/application_1234_0001

Credentials from the standard GCP chain:

Source	How
User credentials	gcloud auth application-default login
Service account key	GOOGLE_APPLICATION_CREDENTIALS=/path/to/key.json
Workload Identity (GKE)	No configuration needed

Azure Data Lake Storage (ADLS Gen2)

pip install -e ".[azure]"
ignis analyze abfs://my-container/spark-logs/application_1234_0001

Credentials from the standard Azure chain:

Source	How
Service principal	AZURE_TENANT_ID + AZURE_CLIENT_ID + AZURE_CLIENT_SECRET
Azure CLI	az login then run ignis normally
Managed identity	No configuration needed

Rules

Rule	What it detects	Default threshold
data-skew	One task takes far longer than its peers in a shuffle stage	max ≥ 5× median task duration
shuffle-size	A stage writes an excessive amount of data to shuffle files	total shuffle write ≥ 1 GB
spill	Tasks spill execution data to disk or show significant memory pressure	any disk spill (WARNING); memory spill ≥ 500 MB (INFO)
partition-count	Shuffle partition count leaves the cluster idle or overwhelms the driver	< 2× executor cores or > 10,000 partitions
failed-tasks	High rate of task failures or speculative task launches in a stage	failure rate ≥ 10% (WARNING); speculation rate ≥ 25% (INFO)
gc-pressure	JVM garbage collection consumes a large fraction of executor run time	GC time ≥ 10% of executor run time (WARNING)

Run ignis rules for a live summary with thresholds.

JSON output

--output json emits a structured document to stdout:

{
  "app_id": "application_1234_0001",
  "app_name": "my-spark-app",
  "finding_count": 1,
  "findings": [
    {
      "rule": "data-skew",
      "severity": "warning",
      "stage_id": 2,
      "stage_name": "groupBy at job.py:42",
      "message": "Stage 2 ('groupBy at job.py:42'): max task 42,300ms vs median 1,800ms (23.5x ratio)",
      "recommendation": "Repartition before the shuffle with a higher partition count, or salt the join/groupBy key to spread work across more tasks."
    }
  ]
}

Development

git clone https://github.com/skatz1990/ignis
cd ignis
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest

Project layout

ignis/
  parser/     NDJSON event log parsing → Application/Stage/Task models
  rules/      Diagnostic rules (one module per rule)
  reporter/   Terminal (rich) and JSON output
  cli.py      Entry point — ignis analyze <path>, ignis rules
tests/
  fixtures/   Hand-crafted NDJSON snippets that trigger each rule
docs/
  rules.md    Detailed explanation of each rule and its detection logic