SKEW

Task skew

One slow task holds a whole stage back while the rest of the cluster waits.

Outcome: Speed
Scope: Judged per stage
Usual fix: A change in the job code
CI gate: Budget flag --max-skew

An example finding

Where you see it

Dashboard

The SKEW tag marks it in the action region of the dashboard, among the widgets that rank a fix. Every affected stage is flagged, not only the worst.How to read it in the findings guide.

Agent (MCP)

An assistant can read this entry with the list_detectors tool and its background with get_finding_documentation. The call and the response below come from a real run of the MCP server.

Call

{
  "tool": "list_detectors",
  "arguments": {
    "type": "skew"
  }
}

Response

{
  "count": 1,
  "detectors": [
    {
      "type": "skew",
      "name": "task skew",
      "tag": "SKEW",
      "tagSharedWith": [],
      "family": "speed",
      "pitch": "One slow task holds a whole stage back while the rest of the cluster waits.",
      "region": "action",
      "scope": "stage",
      "fixEffort": "code",
      "thresholdSummary": "P95 task time over 3× the median (the longest task on stages under 20 tasks), and the slow tasks read correspondingly more data",
      "detectors": [
        {
          "detector": "skew",
          "version": 1,
          "thresholds": {
            "ratioWarn": 3,
            "minTasksForP95": 20,
            "dataShareMin": 0.5,
            "floorPctWarn": 0.005
          }
        }
      ],
      "pages": {
        "catalog": "docs/detectors/skew.html",
        "guide": "docs/user-guide/understanding-findings.html#skew"
      },
      "tuning": {
        "kind": "page",
        "anchor": "#bottleneck-skew",
        "path": "docs/tuning-reference/bottleneck-skew.html#bottleneck-skew"
      },
      "gate": {
        "cliFlag": "--max-skew",
        "airflowParam": "max_skew_ratio"
      }
    }
  ]
}

get_finding_documentation call

{
  "tool": "get_finding_documentation",
  "arguments": {
    "type": "skew"
  }
}

Its response, with each Markdown body replaced by its length

{
  "type": "skew",
  "name": "Task Skew",
  "detectionDoc": {
    "tag": "SKEW",
    "title": "Task skew",
    "content": "<5681 characters of Markdown>"
  },
  "tuningDoc": {
    "anchor": "#bottleneck-skew",
    "title": "Task Skew",
    "content": "<7179 characters of Markdown>"
  }
}

CI (CLI)

List this entry from the command line. The output below is a real run of the command. Budgets and exit codes are in the CI guide.

sparkforensics-analyze detectors --type skew --format md

# Detectors (1)

## Speed

| Tag | Type | Scope | Fix | Gate | What it finds |
| --- | --- | --- | --- | --- | --- |
| SKEW | `skew` | stage | code | `--max-skew` | One slow task holds a whole stage back while the rest of the cluster waits. |

Report only this finding type for a run with --type:

sparkforensics-analyze <event-log> --type skew --impact critical,warning,info

Fail the build when the run breaches a budget for it:

sparkforensics-analyze <event-log> --max-skew <n>

Airflow

The operator takes the budget as the max_skew_ratio parameter.

check_spark_job = SparkForensicsOperator(
    task_id="check_spark_job",
    log_source=...,
    backend=SubprocessAnalyzeHook(),
    max_skew_ratio=...,  # your budget
)

The operator works per run, and its summary counts findings per impact band. See the operator for its parameters.

How it is measured

Thresholds

P95 task time over 3× the median (the longest task on stages under 20 tasks), and the slow tasks read correspondingly more data.

skew version 1

Threshold	Default
`ratioWarn`	3
`minTasksForP95`	20
`dataShareMin`	0.5
`floorPctWarn`	0.005

Each rule's impact bands and fixed fallbacks are in the threshold tables.

Override a default per run with --thresholds; see Tuning detector thresholds.

Docs

This finding has its own tuning page: Spark tuning reference.
Findings guide: SKEW in Understanding findings.