Onboarding my organization#

What information does Union need to set up my service?#

When you initially onboard your organization you must specify which cloud provider(s) you wish to use and the configuration of the machine types you want.

For details, see Configuring your data plane.

How do I change the machine types in my cluster?#

If you have already been onboarded and wish to change your machine types, Union will need to re-configure your node groups (in AWS) or instance groups (in GCP). To initiate the process, submit the Node Group Configuration Change form.

Data storage and handling#

How does Union store my data?#

When data is passed from task to task in a workflow (and output at the end of the workflow), the workflow engine has to store this data so that it can persist after each task container is discarded.

The system distinguishes between metadata and raw data. Primitive values (int, str, etc.) are stored directly in the metadata store while complex data objects (pandas.DataFrame, FlyteFile, etc.) are stored by reference with the reference in metadata and the actual data in the raw data store.

The metadata is stored in your data plane in a preconfigured S3/GCS bucket in your data plane. This data is accessible to the control plane. It is to run and manage workflows and is surfaced in the web UI.

The raw data store is, by default, also located in a pre-configured S3/GCS bucket in your data plane. But this location can be overridden per workflow or per execution using the raw data prefix parameter (see, in this FAQ section, Can I change the raw data storage location?). This allows the actual location to be re-configured by you.

FlyteFile and FlyteDirectory, like all complex data objects, have their underlying data stored in the raw data store (either the OOTB default or whatever you have configured with the raw data prefix), but, in the case of these two classes, you can also specify a custom location on object initialization in your task code (see, in this FAQ section, Where do FlyteFile and FlyteDirectory store their data?).

The data in the raw data store is not accessible to the control plane and will only be surfaced in the web UI unless your code explicitly does so (through, for example, a Flyte Deck).

See also:

Can I change the raw data storage location?#

Yes. There are a number of ways to do this:

  • When registering your workflow:

  • At the execution level:

    • In the web UI, set the Raw output data config parameter in the execution dialog.

These options change the raw data location for all large types (FlyteFile, FlyteDirectory, DataFrame, any other large data object). If you are only concerned with controlling where raw data used by FlyteFile or FlyteDirectory is stored, you can set the remote_path parameter in your task code when initializing objects of those types.

See also:

Can I use my own blob store for data storage that I handle myself?#

Yes. You can certainly configure your own blob storage and then use your chosen library (like boto3, for example) to interact with that storage within your task code. The only caveat is that you must ensure that your task code has access to the storage (see Enabling AWS S3 or Enabling Google Cloud Storage).

Can I control access to my own blob store?#

Yes. As with all resources used by your Flyte task code, the storage must be accessible from within the cluster running your code on your data plane. However, the data plane is your own and you have full control over access (see Enabling AWS S3 or Enabling Google Cloud Storage).

Could someone maliciously delete or otherwise access my raw data?#

No. Your raw data resides in your data plane and is stored either in the default raw data storage or in storage that you set up yourself. In either case, you control access to it.

The Union team does have access to your data plane for purposes of maintenance but does not have access to your raw data, secrets in secret managers, database, etc. unless you choose to permit such access.

Having said that, since the data plane is yours, you are ultimately responsible for preventing access by malicious third parties.

Can I use s3fs from within a task?#

Yes, but you probably don’t need to.

s3fs is a FUSE-based filesystem backed by Amazon S3. It is possible to set up s3fs in your task container image and use it from within your task code.

However, in most cases using either FlyteFile/FlyteDirectory or a library like boto3 to access an S3 bucket directly is preferred (and easier).

If you do need to use s3fs, here are the basic steps:

  • Set up the S3 bucket that you wish to access.

  • Enable access to the bucket from your task code by configuring an appropriate IAM policy. See Enabling AWS S3.

  • Specify your task container image to have s3fs correctly installed and configured.

  • In the task decorator, configure a PodTemplate to run the task container in privileged mode (see links below).

  • In your task code, invoke the s3fs command line tool to mount the S3-backed volume. For example:

subprocess.run(['s3fs', bucket_and_path, mount_point, '-o', 'iam_role=auto'], check=True)

See also:

Can I use BigQuery from within a task?#

If your Union data plane is running on GCP, access to BigQuery should be enabled by default and bound to the default Google Service Account (referred to in this documentation as <UserFlyteGSA>. For details see Enabling GCP resources. If you want to bind it to a different GSA, follow the instructions in Enabling BigQuery.

To actually access your BigQuery instance from your code, you will need to use a BigQueryTask. For details see BigQuery Query.

FlyteFile and FlyteDirectory#

Where do FlyteFile and FlyteDirectory store their data?#

FlyteFile and FlyteDirectory are two Python classes provided by Flyte to make it easy to pass files from one task to the next within a workflow. They do this by wrapping a file or directory location path and, if necessary, maintaining the persistence of the referenced file or directory across task containers.

When you return a FlyteFile (or FlyteDirectory) object from a task, Flyte checks to see if the underlying file or directory is local to the task container or is already remotely located. If it is local then Flyte automatically uploads it to an object store so that it is not lost when the task container is discarded on task completion. If the file is already remote then no upload is performed.

When the FlyteFile (or FlyteDirectory) is passed into the next task, the location of the source file (or directory) is available within the object and it can be downloaded and opened.

By default, when Flyte uploads a local file or directory (as opposed to the case where the source data is already remote), it is stored in the default raw data store (an S3 or GCS bucket) configured in your data plane during setup.

Optionally, you can set up your own bucket and set the raw data prefix parameter to point to it. In that case, Flyte will use this bucket for FlyteFile/FlyteDirectory storage. This setting can be done on registration or per execution on the command line or in the UI.

In either case, the data stored in the bucket is placed in a randomly-named directory with a different random name generated for each FlyteFile (FlyteDirectory) data write. This guarantees that you never overwrite your data.

A further variation is to specify remote_path when initializing your FlyteFile (or FlyteDirectory), in which case the underlying data is written to that precise location with no randomization. In this case, subsequent runs using the same remote_path will overwrite data.

See also:

Can I accidentally overwrite FlyteFile data?#

In general, no. When a task returns a FlyteFile or FlyteDirectory whose source is local to the origin container, Flyte automatically uploads it to a location with a randomized path in the raw data store. This ensures that subsequent runs will not overwrite earlier data.

The only exception is if you explicitly initialize the FlyteFile or FlyteDirectory with a remote_path. In that case, the storage location used is precisely that specified. No randomization occurs so successive runs using the same remote_path will overwrite the same location.

See also:

  • In this FAQ section, Where do FlyteFile and FlyteDirectory store their data?

  • FlyteFile

Can I use my own blob store for FlyteFile and FlyteDirectory data storage?#

Yes. If you do not want to use the default raw output store that is provided with your data plane you can configure your own storage. If you do this, you must ensure that your task code has access to this custom storage (see Enabling AWS S3 or Enabling Google Cloud Storage).

You then have two options for using that storage for FlyteFile and FlyteDirectory:

  • Specify the custom storage location in the output location prefix parameter either on workflow registration or per execution. Your custom storage will be used instead of the default pre-configured raw data store, but the file data will be managed in the same way, using a randomized location within that store for each run, ensuring no overwrites.

  • Specify the remote_path when initializing your FlyteFile or FlyteDirectory object in your task code. The precise location specified will be used with no randomization so avoiding overwrites is up to you.

See also:

  • In this FAQ section, Where do FlyteFile and FlyteDirectory store their data?

  • FlyteFile

How do the typed aliases of FlyteFile and FlyteDirectory work?#

You may notice that flytekit defines some aliases of FlyteFile with specific type annotations. Specifically, FlyteFile has the following aliases for specific file types:

  • HDF5EncodedFile

  • HTMLPage

  • JoblibSerializedFile

  • JPEGImageFile

  • PDFFile

  • PNGImageFile

  • PythonPickledFile

  • PythonNotebook

  • SVGImageFile

Similarly, FlyteDirectory has the following aliases:

  • TensorboardLogs

  • TFRecordsDirectory

These aliases can be used when handling a file or directory of the specified type, however, doing so is entirely optional. Under the covers, the object itself will still be a simple FlyteFile or FlyteDirectory. The aliased version of the classes do not perform any checks on the actual content of the file, they are simply syntactic markers that enforce agreement between type annotations in the signatures of task functions.

Building and running workflows#

What SDK should I download and use in workflow code?#

You should install the unionai SDK, which will install the unionai and flytekit SDKs and the unionai command-line tool. You will need to use the flytekit SDK the majority of the time in the code to import core features and use the unionai SDK for Union-specific features, such as Artifacts.

To install the unionai SDK, flytekit SDK, and unionai CLI, run the following command:

pip install unionai

How do I authenticate uctl and unionai to Union?#

The command-line tools uctl and unionai need to authenticate in order to connect with your Union instance (for example, when registering a workflow). There are three ways to set up authentication.

  1. PKCE: This is the default method. When using this method, a browser pops up to authenticate the user.

  2. DeviceFlow: A URL will be output to your terminal. Navigate to it in your browser and follow the directions.

  3. ClientSecret: This is the headless option. It can be used, for example, by CI bots. With this method, you create a Union application and configure your tools to pass the Client ID and App Secret to Union.

These methods are all configured in the config.yaml that your uctl or unionai command uses. See cli-authentication.md for full details.

How do I specify resource requirements for a task?#

You can specify either requests or limits (or both) on the resources that will be used by a specific task when it runs in its container. This is done by setting the requests or limits property in the @task decorator to a Resources configuration object. Within the Resources object you can specify the number of CPU cores, the number of GPU cores, the amount of main memory, the amount of persistent storage, and the amount of ephemeral storage.

You can also override the settings in the @task in a for more fine-grained control using the with_overrides method when invoking the task function.

See also:

What command-line tools should I use to register and run workflows?#

You should use the unionai CLI to register and run workflows and perform other operations on the command line. The unionai CLI is installed when you pip install unionai, which will also install the unionai SDK and flytekit SDK.

To install the unionai SDK, flytekit SDK, and unionai CLI, run the following command:

pip install unionai

What happens if an automated process launches a very large number of workflows?#

By default, Union has a built-in limiting mechanism that prevents more than 10,000 concurrent workflow executions per data plane cluster (equivalently, per organization). This limit can be adjusted on a per-customer basis (talk to the Union team).

Executions beyond the limit will be executed as soon as resources become available. While waiting, the workflow execution will be reported as in the UNKNOWN state.

This limit prevents workflow requests from overwhelming the cluster and, in effect, performing a self-caused denial of service attack.

How can I constrain the number of parallel executions for large, complex workflows?#

Workflows can quickly get complex when dynamic workflows iterate over varying length inputs, workflows call subworkflows, and map tasks iterate over a series of inputs. There are two levers to control the parallelism of a workflow: max_parallelism which controls parallelism at a workflow level, and concurrency which controls parallelism at a map task level. Another way of thinking about this is that max_parallelism controls the number of simultaneous executions of all tasks except for map tasks which are controlled separately. This means that the total number of simultaneous executions during a workflow run cannot exceed max_parallelism * concurrency which would be the case if each parallel execution at the workflow level had its own map task.

By default max_parallelism is set to 25. If concurrency is not set for a map task, the current default behavior is to execute over all inputs to the map task. The trade-off that must be balanced when setting max_parallelism and concurrency is with resource availability at a workflow level. If parallelism is too high, tasks can time out before resources can be allocated to them, making it important to consider the resource requirements of your tasks that will run in parallel.

When interpreting parallelism in the UI, it is important to note that dynamic workflows will immediately list all planned executions, even if the number exceeds max_parallelism. However, this does not mean that all the executions are running. By toggling any embedded tasks or subworkflows, you should see an UNKNOWN status for any tasks that have not yet been processes due to the limitations of max_parallelism.

See also: