Scalable Job Tracking

Launch this tutorial in a Jupyter Notebook on Binder:

The Python bindings provide two scalable mechanisms for tracking jobs:

  • Poll-based tracking: The Schedd can be periodically polled through the use of Schedd.xquery to get job status information.
  • Event-based tracking: Using the job's user log, Python can see all job events and keep an in-memory representation of the job status.

Both poll- and event-based tracking have their strengths and weaknesses; the intrepid user can even combine both methodologies to have extremely reliable, low-latency job status tracking.

In this module, we outline the important design considerations behind each approach and walk through examples.

Poll-based Tracking

Poll-based tracking involves periodically querying the schedd(s) for jobs of interest. We have covered the technical aspects of querying the Schedd in prior tutorials. Beside the technical means of polling, important aspects to consider are how often the poll should be performed and how much data should be retrieved.

Note: When Schedd.xquery is used, the query will cause the schedd to fork up to SCHEDD_QUERY_WORKERS simultaneous workers. Beyond that point, queries will be handled in a non-blocking manner inside the main condor_schedd process. Thus, the memory used by many concurrent queries can be reduced by decreasing SCHEDD_QUERY_WORKERS.

A job tracking system should not query the Schedd more than once a minute. Aim to minimize the data returned from the query through the use of the projection; minimize the number of jobs returned by using a query constraint. Better yet, use the AutoCluster flag to have Schedd.xquery return a list of job summaries instead of individual jobs.

Advantages:

  • A single entity can poll all condor_schedd instances in a pool; using htcondor.poll, multiple Schedds can be queried simultaneously.
  • The tracking is resilient to bugs or crashes. All tracked state is replaced at the next polling cycle.

Disadvantages:

  • The amount of work to do is a function of the number of jobs in the schedd; may scale poorly once more than 100,000 simultaneous jobs are tracked.
  • Each job state transition is not seen; only snapshots of the queue in time.
  • If a job disappears from the Schedd, it may be difficult to determine why (Did it finish? Was it removed?)
  • Only useful for tracking jobs at the minute-level granularity.

Event-based Tracking

Each job in the Schedd can specify the UserLog attribute; the Schedd will atomically append a machine-parseable event to the specified file for every state transition the job goes through. By keeping track of the events in the logs, we can build an in-memory representation of the job queue state.

Advantages:

  • No interaction with the condor_schedd process is needed to read the event logs; the job tracking effectively places no burden on the Schedd.
  • In most cases, the Schedd writes to the log synchronously after the event occurs. Hence, the latency of receiving an update can be sub-second.
  • The job tracking scales as a function of the event rate, not the total number of jobs.
  • Each job state is seen, even after the job has left the queue.

Disadvantages:

  • Only the local condor_schedd can be tracked; there is no mechanism to receive the event log remotely.
  • Log files must be processed from the beginning, with no rotations or truncations possible. Large files can take a large amount of CPU time to process.
  • If every job writes to a separate log file, the job tracking software may have to keep an enormous number of open file descriptors. If every job writes to the same log file, the log file may grow to many gigabytes.
  • If the job tracking software misses an event (or an unknown bug causes the condor_schedd to fail to write the event), then the job tracker may believe a job incorrectly is stuck in the wrong state.

At a technical level, event tracking is implemented with the htcondor.JobEventLog class.

>>> jel = htcondor.JobEventLog("/tmp/job_one.log")
>>> for event in jel.events(stop_after=0):
...     print event

The return value of JobEventLog.events() is an iterator over htcondor.JobEvent objects. The example above does not block.