A main topic of a workflow engine is the exception handling of tasks which failed during batch processing.
Luigi does not provide a pre-defined strategy for exception handling which have to be implemented individually
by the user, if possible. However, Luigi offers a mechanism for so called 'idempotent' tasks which are only run
once, given a specific set of parameters. Hereby, Luigi skipps complete tasks while running a pipeline
the second time after encounting an unresolvable dependency (i.e. a predecessor task which failed while execution).
A more detailed strategy for exception handling and the implementation of idempotent tasks are given in the following sections.
Exception: Workflow Engine
Description:
Exception handling through the workflow engine (e.g. after power failure, ...)
Approaches:
| ID | Process | Exception Handling |
|---|---|---|
| 1 | None. | None. |
| 2 | Tasks are stored within a database and updated whenever a task terminates.
| While start up Luigi, check and re-run open pipeline tasks. |
A successful task (with a docker container as target) can be defined in different ways:
- Given, a result file with a specified format (Drawback: additional effort for data handling, error-prone due file formatting).
- Exit code of the docker container is equal to 0 ( = successful).
The second approach is more accurate due simplicity.
Exception: Algorithm
Description:
Exception handling through algorithm containers (e.g. invalid parameters, unreachable database ...)
Approaches:
| ID | Process | Exception Handling |
|---|---|---|
| 1 | None. | None. |
| 2 | An algorithm continuously saves results which are deleted if an exception is
| While encountering an exception, clean up inconsistent |
| 3 | An algorithm persistently saves all or none results.
| Results are only saved if no exception was throws. |
Note 1:
Approaches (2) and (3) may result in inconsistent results if an algorithm saves result fragments and terminates immediately during a
superior exception (e.g. power failure). Therefore, each algorithm is responsible for saving consistent results.
Note 2:
We do not provide any rollback meachanism. Inadequate results are not detected nor cleaned up by Luigi. Therefore, consistent results
are crucial. (Advantages: Reduce complexity of workflow engine and hereby the error rate due rollbacks).
It is also possible to save (consistent) result fragments iteratively. However, if the algorithm is restarted after a failure (e.g. power failure) it has to
be able to handle such fragments individually.
Idempotent Algorithm
Description:
Avoiding to run an algorithm twice, which results are already available.
Approaches:
| ID | Process |
|---|---|
| 1 | Using Luigi's mechanism of idempotent algorithms (see preliminary). |
| 2 | Each algorithm is responsible for fulfilling the idempotent property by itself. Luigi's mechanism is omitted. Drawback: More complexer algorithms, error-prone due wrong implementations |
Given the mechanism of idempotent tasks and the fact as we can verify an algorithm's exit code, Luigi could simply infer if an algorithm has to be
run or was already completed. A problem appears after restarting Luigi which discards all containers and there exit codes.
A recommended solution is to save all container's exit codes and standard streams (stdout/strerr) within Luigi's database. The solution supports
traceability for historical batch processes and provides Luigi a way to identify complete tasks.