convalesce-emit-pyspark

Report a PySpark driver to Convalesce: the script it ran, what it was called with, and whether it failed in Python.

A driver can fail before Spark runs anything, say by reading a table that is not there. Spark still ends the application normally, so the listener that convalesce-emit-spark installs reports a success. This package hooks the driver's own Python and sends one driver_failure observation, which marks the application's run failed, with the Python exception and its traceback.

Whether it failed or not, it also sends one driver_script observation as the driver exits: the script's text and the arguments it was given. What ran is the file on the machine that ran it, which a repository can only suggest.

Install

Into the Python the driver runs on:

pip install convalesce-emit-pyspark

Turn it on

With no code change, set this where spark-submit runs:

export CONVALESCE_PYSPARK_DRIVER_HOOK=true

The package ships a .pth file, which Python runs at start-up. It reads that one variable and does nothing else unless it is true, so installing the package changes nothing until you set it.

Or in the driver, before anything can fail:

import convalesce_emit_pyspark

convalesce_emit_pyspark.install()

Either way, set CONVALESCE_INGEST_KEY too; without it nothing is hooked. Run it beside convalesce-emit-spark, which reports the application itself.

What is reported

driver_script, once per driver

Sent as the interpreter exits, after a run that succeeded and after one that failed, for a driver that started a Spark application:

• the application's id and name
• argv: the script and its arguments, with credentials masked (--password hunter2 and --url=postgresql://etl:hunter2@db/shop both lose the password)
• source: the script's path, its text with credentials masked the same way, and the sha256 of the file as it is on the driver's machine. Text over 60,000 characters is cut from the end and marked truncated; the hash is always of the whole file
• the Python and PySpark versions

Two settings, both on unless set to false, 0, no or off:

Variable	Off means
CONVALESCE_SEND_SOURCE	no source
CONVALESCE_SEND_ARGUMENTS	argv holds the script's path only, here and in driver_failure

driver_failure, at most once per driver

For whichever comes first:

• an exception nobody caught, in the main thread or another thread
• sys.exit(n) with n not zero, sent as a SystemExit

It carries the application's id and name, the exception (type, message, traceback, and its cause), argv as above, and the Python and PySpark versions. A notebook sends no argv: its command line is the kernel's.

Which run of the platform, on both

• databricks, on Databricks: the job run's id, the task run's id, the job's id and name, the notebook's path, the cluster's id, and the workspace's id and URL, each one that is known. The job run's id puts a failure on the right job run of a cluster that several job runs share. See What is read on Databricks.
• attempt, in a YARN container: which try of the application the driver is, from 1. YARN runs a failed cluster-mode driver again under the same application id, and each try reports.

Both observations are sent before the driver exits, which delays that exit by at most CONVALESCE_TIMEOUT per attempt. The hook never raises and never changes the exit code; the exception still prints as it would have.

A driver usually stops its SparkSession before it exits, and a stopped session no longer names its application. The hook wraps SparkContext.stop to read the application's id and name first, and changes nothing else about it.

Limits:

• raise SystemExit(n) is not seen, only sys.exit(n)
• a sys.exit(n) whose SystemExit the driver catches and swallows is still reported
• source is the script the driver was started with (spark-submit job.py), not the modules it imports. Where a platform's launcher runs the script (AWS Glue's runscript.py), source is the script and never the launcher

Notebooks

A notebook reports a failure. The cell that raises sends driver_failure, and on Spark Connect (Databricks serverless) the run's start before it and its end after it, all three with the same databricks object. A notebook has no script, so it sends no driver_script and no argv, and a cell that passes sends nothing and ends nothing: a notebook that never fails sends nothing.

On Spark Connect the run always has a name. It is the first of these that is known:

1. the script's file name, without .py, for a Python file task, which sends driver_script as it always did
2. the last part of the notebook's path
3. the Databricks job's name
4. the session's spark.app.name
5. notebook inside a notebook kernel, pyspark anywhere else

What is read on Databricks

Two places are asked, the first one first:

1. The session's configuration, by setting name: spark.databricks.job.runId, spark.databricks.job.id, spark.databricks.notebook.path, spark.databricks.clusterUsageTags.clusterId, spark.databricks.clusterUsageTags.clusterOwnerOrgId and spark.databricks.workspaceUrl. A classic cluster answers all of them. A serverless session answers for the cluster and the workspace's URL.
2. The run's context, which the dbutils of a notebook or a Python file task hands out. It fills in what the configuration did not say. These attributes are read from it by name:

Attribute	Sent as
multitaskParentRunId	databricks.job_run_id, the job run
currentRunId	databricks.task_run_id, the task's run; also job_run_id when the context names no job run
jobId	databricks.job_id
jobName	databricks.job_name
notebook_path	databricks.notebook_path, for a notebook task
clusterId	databricks.cluster_id
orgId	databricks.workspace_id
jobTaskType	not sent; it says whether the task is a notebook

Databricks runs a failed task again as a new task run of the same job run, so the two tries share job_run_id and differ in task_run_id.

The context also holds the run's own API token, the API's URL and the user it runs as. The hook takes the attributes in the table out of the context by name and keeps nothing else: no other attribute is stored, logged or sent. The context is asked once per driver, the hook imports no Databricks module to ask it, and a context that will not answer changes nothing.

A driver with no Spark session

A cluster-mode driver on YARN can fail before it starts a Spark session. YARN has made the application and no listener is running to report it, so the hook reports the whole run: SparkListenerApplicationStart (named by the script's file name, timed from when the hook went in), driver_failure, driver_script, then SparkListenerApplicationEnd, each with attempt. Once a driver has asked for a Spark context the listener reports the application and the hook sends only its own two observations.

The ingest key

The key is sent in the Authorization header only. Its value is masked (***) anywhere else it turns up, such as a script with the key written into it, and the observation's excluded says so ({"path": "$", "reason": "ingest key masked"}). A key under 8 characters is too short to look for.

Supported

PySpark 3.3 to 4.0, on Python 3.9 and later.

Configure

Read from the environment: see convalesce-emit.

On AWS Glue, give every setting with CUSTOMER_ in front, in the job parameter --customer-driver-env-vars: CUSTOMER_CONVALESCE_INGEST_KEY=...,CUSTOMER_CONVALESCE_PYSPARK_DRIVER_HOOK=true. Glue installs --additional-python-modules where Python runs the .pth file, so that last one turns the hook on with no code change. driver_script carries the job's script either way, whether the variable or a call to install() turned the hook on.