Job failures can occur for a variety of reasons, but they generally categorize into 3 failure types:
- User-code-related: An error in the user code running inside the container or the json pipeline config.
- Data-related: A problem with the input data such as incorrect file type or file name.
- System- or infrastructure-related: An error in Pachyderm or Kubernetes such as missing credentials, transient network errors, or resource constraints (for example, out-of-memory--OOM--killed).
In this document, we'll show you the tools for determining what kind of failure it is. For each of the failure modes, we’ll describe Pachyderm’s and Kubernetes’s specific retry and error-reporting behaviors as well as typical user triaging methodologies.
Failed jobs in a pipeline will propagate information to downstream pipelines with empty commits to preserve provenance and make tracing the failed job easier. A failed job is no longer running.
In this document, we'll describe what you'll see, how Pachyderm will respond, and techniques for triaging each of those three categories of failure.
At the bottom of the document, we'll provide specific troubleshooting steps for specific scenarios.
Determining the kind of failure¶
First off, you can see the status of Pachyderm's jobs with
pachctl list job, which will show you the status of all jobs. For a failed job, use
pachctl inspect job <job-id> to find out more about the failure. The different categories of failures are addressed below.
User Code Failures¶
When there’s an error in user code, the typical error message you’ll see is
failed to process datum <UUID> with error: <user code error>
This means pachyderm successfully got to the point where it was running user code, but that code exited with a non-zero error code. If any datum in a pipeline fails, the entire job will be marked as failed, but datums that did not fail will not need to be reprocessed on future jobs. You can use
pachctl inspect datum <job-id> <datum-id> or
pachctl logs with the
--datum flags to get more details.
There are some cases where users may want mark a datum as successful even for a non-zero error code by setting the
transform.accept_return_code field in the pipeline config .
Pachyderm will automatically retry user code three (3) times before marking the datum as failed. This mitigates datums failing for transient connection reasons.
pachctl logs --job=<job_ID> or
pachctl logs --pipeline=<pipeline_name> will print out any logs from your user code to help you triage the issue. Kubernetes will rotate logs occasionally so if nothing is being returned, you’ll need to make sure that you have a persistent log collection tool running in your cluster. If you set
enable_stats:true in your pachyderm pipeline, pachyderm will persist the user logs for you.
In cases where user code is failing, changes first need to be made to the code and followed by updating the pachyderm pipeline. This involves building a new docker container with the corrected code, modifying the pachyderm pipeline config to use the new image, and then calling
pachctl update pipeline -f updated_pipeline_config.json. Depending on the issue/error, user may or may not want to also include the
--reprocess flag with
When there’s an error in the data, this will typically manifest in a user code error such as
failed to process datum <UUID> with error: <user code error>
This means pachyderm successfully got to the point where it was running user code, but that code exited with a non-zero error code, usually due to being unable to find a file or a path, a misformatted file, or incorrect fields/data within a file. If any datum in a pipeline fails, the entire job will be marked as failed. Datums that did not fail will not need to be reprocessed on future jobs.
Just like with user code failures, Pachyderm will automatically retry running a datum 3 times before marking the datum as failed. This mitigates datums failing for transient connection reasons.
Data failures can be triaged in a few different way depending on the nature of the failure and design of the pipeline.
In some cases, where malformed datums are expected to happen occasionally, they can be “swallowed” (e.g. marked as successful using
transform.accept_return_codes or written out to a “failed_datums” directory and handled within user code). This would simply require the necessary updates to the user code and pipeline config as described above. For cases where your code detects bad input data, a "dead letter queue" design pattern may be needed. Many pachyderm developers use a special directory in each output repo for "bad data" and pipelines with globs for detecting bad data direct that data for automated and manual intervention.
Pachyderm's engineering team is working on changes to the Pachyderm Pipeline System in a future release that may make implementation of design patterns like this easier. Take a look at the pipeline design changes for pachyderm 1.9
If a few files as part of the input commit are causing the failure, they can simply be removed from the HEAD commit with
finish commit. The files can also be corrected in this manner as well. This method is similar to a revert in Git -- the “bad” data will still live in the older commits in Pachyderm, but will not be part of the HEAD commit and therefore not processed by the pipeline.
If the entire commit is bad and you just want to remove it forever as if it never happened,
delete commit will both remove that commit and all downstream commits and jobs that were created as downstream effects of that input data.
System-level failures are the most varied and often hardest to debug. We’ll outline a few common patterns and triage steps. Generally, you’ll need to look at deeper logs to find these errors using
pachctl logs --pipeline=<pipeline_name> --raw and/or
kubectl logs pod <pod_name>.
Here are some of the most common system-level failures:
- Malformed or missing credentials such that a pipeline cannot connect to object storage, registry, or other external service. In the best case, you’ll see
permission deniederrors, but in some cases you’ll only see “does not exist” errors (this is common reading from object stores)
- Out-of-memory (OOM) killed or other resource constraint issues such as not being able to schedule pods on available cluster resources.
- Network issues trying to connect Pachd, etcd, or other internal or external resources
- Failure to find or pull a docker image from the registry
For system-level failures, Pachyderm or Kubernetes will generally continually retry the operation with exponential backoff. If a job is stuck in a given state (e.g. starting, merging) or a pod is in
CrashLoopBackoff, those are common signs of a system-level failure mode.
Triaging system failures varies as widely as the issues do themselves. Here are options for the common issues mentioned previously. - Credentials: check your secrets in k8s, make sure they’re added correctly to the pipeline config, and double check your roles/perms within the cluster - OOM: Increase the memory limit/request or node size for your pipeline. If you are very resource constrained, making your datums smaller to require less resources may be necessary. - Network: Check to make sure etcd and pachd are up and running, that k8s DNS is correctly configured for pods to resolve each other and outside resources, firewalls and other networking configurations allow k8s components to reach each other, and ingress controllers are configured correctly - Check your container image name in the pipeline config and image_pull_secret.
All your pods or jobs get evicted¶
$ kubectl get all
shows a bunch of pods that are marked
Evicted. If you
kubectl describe ... one of those evicted pods, you see an error saying that it was evicted due to disk pressure.
Your nodes are not configured with a big enough root volume size. You need to make sure that each node's root volume is big enough to store the biggest datum you expect to process anywhere on your DAG plus the size of the output files that will be written for that datum.
Let's say you have a repo with 100 folders. You have a single pipeline with this repo as an input, and the glob pattern is
/*. That means each folder will be processed as a single datum. If the biggest folder is 50GB and your pipeline's output is about 3 times as big, then your root volume size needs to be bigger than:
50 GB (to accommodate the input) + 50 GB x 3 (to accommodate the output) = 200GB
In this case we would recommend 250GB to be safe. If your root volume size is less than 50GB (many defaults are 20GB), this pipeline will fail when downloading the input. The pod may get evicted and rescheduled to a different node, where the same thing will happen.
Pipeline exists but never runs¶
You can see the pipeline via:
$ pachctl list pipeline
But if you look at the job via:
$ pachctl list job
It's marked as running with
0/0 datums having been processed. If you inspect the job via:
$ pachctl inspect job
You don't see any worker set. E.g:
Worker Status: WORKER JOB DATUM STARTED ...
If you do
kubectl get pod you see the worker pod for your pipeline, e.g:
But it's state is
First make sure that there is no parent job still running. Do
pachctl list job | grep yourPipelineName to see if there are pending jobs on this pipeline that were kicked off prior to your job. A parent job is the job that corresponds to the parent output commit of this pipeline. A job will block until all parent jobs complete.
If there are no parent jobs that are still running, then continue debugging:
Describe the pod via:
$kubectl describe po/pipeline-foo-5-v1-273zc
If the state is
CrashLoopBackoff, you're looking for a descriptive error message. One such cause for this behavior might be if you specified an image for your pipeline that does not exist.
If the state is
Pending it's likely the cluster doesn't have enough resources. In this case, you'll see a
could not schedule type of error message which should describe which resource you're low on. This is more likely to happen if you've set resource requests (cpu/mem/gpu) for your pipelines. In this case, you'll just need to scale up your resources. If you deployed using
kops, you'll want to do edit the instance group, e.g.
kops edit ig nodes ... and up the number of nodes. If you didn't use
kops to deploy, you can use your cloud provider's auto scaling groups to increase the size of your instance group. Either way, it can take up to 10 minutes for the changes to go into effect.
For more information, see Autoscale Your Cluster.