Dagster Instance#

Overview#

The DagsterInstance defines all of the configuration that Dagster needs for a single deployment - for example, where to store the history of past runs and their associated logs, where to stream the raw logs from op compute functions, and how to launch new runs.

All of the processes and services that make up your Dagster deployment should share a single instance config file so that they can effectively share information.

Some important configuration, like execution parallelism, is set on a per-job basis rather than on the instance.

Default local behavior#

When you launch a Dagster process, like Dagit or the Dagster CLI commands, Dagster attempts to load your instance. If the environment variable DAGSTER_HOME is set, Dagster will look for an instance config file at $DAGSTER_HOME/dagster.yaml. This file contains each of the configuration settings that make up the instance.

By default (if dagster.yaml is not present or nothing is specified in that file), Dagster will store this information on the local filesystem, laid out like this:

$DAGSTER_HOME
├── dagster.yaml
├── history
│   ├── runs
│   │   ├── 00636713-98a9-461c-a9ac-d049407059cd.db
│   │   └── ...
│   └── runs.db
└── storage
    ├── 00636713-98a9-461c-a9ac-d049407059cd
    │   └── compute_logs
    │       ├── my_solid.compute.complete
    │       ├── my_solid.compute.err
    │       ├── my_solid.compute.out
    │       └── ...
    └── ...

The runs.db and {run_id}.db files are SQLite database files recording information about runs and per-run event logs respectively. The compute_logs directories (one per run) contain the stdout and stderr logs from the execution of the compute functions of each op.

If DAGSTER_HOME is not set, the Dagster tools will use an ephemeral instance for execution. In this case, the run and event log storages will be in-memory rather than persisted to disk, and filesystem storage will use a temporary directory that is cleaned up when the process exits. This is useful for tests and is the default when invoking Python APIs such as JobDefinition.execute_in_process directly.

Instance Configuration YAML#

In persistent Dagster deployments, you will typically want to configure many of the components on the instance. For example, you may want to use a Postgres instance to store runs and the corresponding event logs, and to stream compute logs to an S3 bucket.

To do this, provide a $DAGSTER_HOME/dagster.yaml file. Dagit and all Dagster tools will look for this file on startup. In the dagster.yaml file, you can configure many different aspects of your Dagster Instance, all of which are detailed below.

Note that Dagster supports retrieving instance YAML values from environment variables, using an env: key instead of a string literal value. Examples of using env: are included in the sample configurations below.

Dagster Storage#

Dagster storage configures how job and asset history is persisted - this includes metadata on runs, event logs, schedule/sensor ticks, and other useful data.

To configure storage, you should set the storage attribute in your dagster.yaml. There are three available options:

Sqlite storage (Default)#

DagsterSqliteStorage uses a Sqlite DB for storage.
# there are two ways to set storage to SqliteStorage

# this config manually sets the directory (`base_dir`) for Sqlite to store data in:
storage:
  sqlite:
    base_dir: /path/to/dir

# and this config grabs the directory from an environment variable
storage:
  sqlite:
    base_dir:
      env: SQLITE_STORAGE_BASE_DIR

Postgres storage#

DagsterPostgresStorage uses a Postgres DB as the backing storage solution. This requires that the `dagster-postgres` library be installed.
# Postgres storage can be set using either credentials or a connection string.  This requires that
# the `dagster-postgres` library be installed.

# this config manually sets the Postgres credentials
storage:
  postgres:
    postgres_db:
      username: { DAGSTER_PG_USERNAME }
      password: { DAGSTER_PG_PASSWORD }
      hostname: { DAGSTER_PG_HOSTNAME }
      db_name: { DAGSTER_PG_DB }
      port: 5432

# and this config grabs the database credentials from environment variables
storage:
  postgres:
    postgres_db:
      username:
        env: DAGSTER_PG_USERNAME
      password:
        env: DAGSTER_PG_PASSWORD
      hostname:
        env: DAGSTER_PG_HOST
      db_name:
        env: DAGSTER_PG_DB
      port: 5432

# and this config sets the credentials via DB connection string / url:
storage:
  postgres:
    postgres_url: { PG_DB_CONN_STRING }

# This config gets the DB connection string / url via environment variables:
storage:
  postgres:
    postgres_url:
      env: PG_DB_CONN_STRING

MySQL storage#

DagsterMySQLStorage uses a MySQL DB as the backing storage solution. This requires that the `dagster-mysql` library be installed.
# MySQL storage can be set using either credentials or a connection string.  This requires that the
# `dagster-mysql` library be installed.

# this config manually sets the MySQL credentials
storage:
  mysql:
    mysql_db:
      username: { DAGSTER_MYSQL_USERNAME }
      password: { DAGSTER_MYSQL_PASSWORD }
      hostname: { DAGSTER_MYSQL_HOSTNAME }
      db_name: { DAGSTER_MYSQL_DB }
      port: 3306


# and this config grabs the database credentials from environment variables
storage:
  mysql:
    mysql_db:
      username:
        env: DAGSTER_MYSQL_USERNAME
      password:
        env: DAGSTER_MYSQL_PASSWORD
      hostname:
        env: DAGSTER_MYSQL_HOSTNAME
      db_name:
        env: DAGSTER_MYSQL_DB
      port: 3306

# and this config sets the credentials via DB connection string / url:
storage:
  mysql:
    mysql_url: { MYSQL_DB_CONN_STRING }

# this config grabs the MySQL connection string from environment variables
storage:
  mysql:
    mysql_url:
      env: MYSQL_DB_CONN_STRING

Run Launcher#

The run launcher determines where runs are executed.

There are several Dagster-provided options for the Run Launcher; users also can write custom run launchers. See the Run Launcher docs for more information.

To configure the Run Launcher, set run_launcher in your dagster.yaml in one of the following ways:

DefaultRunLauncher (Default)#

The DefaultRunLauncher spawns a new process in the same node as a job's repository location. Please see the Run Launcher docs for deployment information.

run_launcher:
  module: dagster.core.launcher
  class: DefaultRunLauncher

DockerRunLauncher#

The DockerRunLauncher allocates a Docker container per run. Please see the Run Launcher docs for deployment information.

run_launcher:
  module: dagster_docker
  class: DockerRunLauncher

K8sRunLauncher#

The K8sRunLauncher allocates a Kubernetes Job per run. Please see the Run Launcher docs for deployment information.

# there are multiple ways to configure the K8sRunLauncher

# you can set the follow configuration values directly
run_launcher:
  module: dagster_k8s.launcher
  class: K8sRunLauncher
  config:
    service_account_name: pipeline_run_service_account
    job_image: my_project/dagster_image:latest
    instance_config_map: dagster-instance
    postgres_password_secret: dagster-postgresql-secret

# alternatively, you can grab any of these config values from environment variables:
run_launcher:
  module: dagster_k8s.launcher
  class: K8sRunLauncher
  config:
    service_account_name:
      env: PIPELINE_RUN_SERVICE_ACCOUNT
    job_image:
      env: DAGSTER_IMAGE_NAME
    instance_config_map:
      env: DAGSTER_INSTANCE_CONFIG_MAP
    postgres_password_secret:
      env: DAGSTER_POSTGRES_SECRET

Run Coordinator#

The run coordinator determines the policy used to determine the prioritization rules and concurrency limits for runs. Please see the Run Coordinator Docs for more information and for troubleshooting help.

To configure the Run Coordinator, set the run_coodinator key in your dagster.yaml. There are two options:

DefaultRunCoordinator (Default)#

The DefaultRunCoordinator immediately sends runs to the run launcher (no notion of Queued runs).

See the Run Coordinator docs for more information.

# Since DefaultRunCoordinator is the default option, omitting the `run_coordinator` key will also suffice,
# but if you would like to set it explicitly:
run_coordinator:
  module: dagster.core.run_coordinator
  class: DefaultRunCoordinator

QueuedRunCoordinator#

The QueuedRunCoordinator allows you to set limits on the number of runs that can be executing at once. Note that this requires a dagster-daemon process to be active to actually launch the runs.

This run coordinator has several configuration options, which allow for both limiting the overall number of concurrent runs as well as more specific limits based on run tags - for example, you can configure a limit on the number of runs that interact with a particular cloud service that can run concurrently to avoid being throttled.

For more information, see the Run Coordinator docs.

# There are a few ways to configure the QueuedRunCoordinator:

# this first option has concurrency limits set to default values
run_coordinator:
  module: dagster.core.run_coordinator
  class: QueuedRunCoordinator

# this second option manually specifies limits:
run_coordinator:
  module: dagster.core.run_coordinator
  class: QueuedRunCoordinator
  config:
    max_concurrent_runs: 25
    tag_concurrency_limits:
      - key: "database"
        value: "redshift"
        limit: 4
      - key: "dagster/backfill"
        limit: 10

# as always, some or all of these values can be obtained from environment variables:
run_coordinator:
  module: dagster.core.run_coordinator
  class: QueuedRunCoordinator
  config:
    max_concurrent_runs:
      env: DAGSTER_OVERALL_CONCURRENCY_LIMIT
    tag_concurrency_limits:
      - key: "database"
        value: "redshift"
        limit:
          env: DAGSTER_REDSHIFT_CONCURRENCY_LIMIT
      - key: "dagster/backfill"
        limit:
          env: DAGSTER_BACKFILL_CONCURRENCY_LIMIT

Compute Log Storage#

Compute log storage controls the capture and persistence of raw stdout & stderr text logs.

To configure Compute Log Storage, set the compute_logs key in your dagster.yaml.

LocalComputeLogManager (Default)#

LocalComputeLogManager writes stdout & stderr logs to disk.
# there are two ways to set the directory that the LocalComputeLogManager writes
# stdout & stderr logs to

# You could directly set the `base_dir` key
compute_logs:
  module: dagster.core.storage.local_compute_log_manager
  class: LocalComputeLogManager
  config:
    base_dir: /path/to/directory

# Alternatively, you could set the `base_dir` key to an environment variable
compute_logs:
  module: dagster.core.storage.local_compute_log_manager
  class: LocalComputeLogManager
  config:
    base_dir:
      env: LOCAL_COMPUTE_LOG_MANAGER_DIRECTORY

AzureBlobComputeLogManager#

AzureBlobComputeLogManager writes stdout & stderr logs to Azure Blob Storage.
# there are multiple ways to configure the AzureBlobComputeLogManager

# you can set the necessary configuration values directly:
compute_logs:
  module: dagster_azure.blob.compute_log_manager
  class: AzureBlobComputeLogManager
  config:
    storage_account: mycorp-dagster
    container: compute-logs
    secret_key: foo
    local_dir: /tmp/bar
    prefix: dagster-test-

# alternatively, you can obtain any of these config values from environment variables
compute_logs:
  module: dagster_azure.blob.compute_log_manager
  class: AzureBlobComputeLogManager
  config:
    storage_account:
      env: MYCORP_DAGSTER_STORAGE_ACCOUNT_NAME
    container:
      env: CONTAINER_NAME
    secret_key:
      env: SECRET_KEY
    local_dir:
      env: LOCAL_DIR_PATH
    prefix:
      env: DAGSTER_COMPUTE_LOG_PREFIX

S3ComputeLogManager#

S3ComputeLogManager writes stdout & stderr logs to AWS S3.
# there are multiple ways to configure the S3ComputeLogManager

# you can set the config values directly:
compute_logs:
  module: dagster_aws.s3.compute_log_manager
  class: S3ComputeLogManager
  config:
    bucket: "mycorp-dagster-compute-logs"
    prefix: "dagster-test-"

# or grab some or all of them from environment variables
compute_logs:
  module: dagster_aws.s3.compute_log_manager
  class: S3ComputeLogManager
  config:
    bucket:
      env: MYCORP_DAGSTER_COMPUTE_LOGS_BUCKET
    prefix:
      env: DAGSTER_COMPUTE_LOG_PREFIX

Local Artifact Storage#

Local artifact storage is used to configure storage for any artifacts that require a local disk, or when using the filesystem IO manager to store inputs and outputs. See IO Managers for more information on how other IO managers store artifacts.

To configure Local Artifact Storage, set local_artifact_storage as follows in your dagster.yaml:

LocalArtifactStorage (Default)#

LocalArtifactStorage is currently the only option for Local Artifact Storage. This configures the directory used by the default filesystem IO Manager, as well as any artifacts that require a local disk.
# there are two possible ways to configure LocalArtifactStorage

# example local_artifact_storage setup pointing to /var/shared/dagster directory
local_artifact_storage:
  module: dagster.core.storage.root
  class: LocalArtifactStorage
  config:
    base_dir: "/path/to/dir"

# alternatively, `base_dir` can be set to an environment variable
local_artifact_storage:
  module: dagster.core.storage.root
  class: LocalArtifactStorage
  config:
    base_dir:
      env: DAGSTER_LOCAL_ARTIFACT_STORAGE_DIR

Telemetry#

This allows opting in/out (set to true by default) of Dagster collecting anonymized usage statistics.

To configure Telemetry, set the telemetry key in your dagster.yaml.

For more information on how and why we use telemetry, please visit the Telemetry Docs.

# Allows opting out of Dagster collecting usage statistics.
telemetry:
  enabled: false

Code servers#

The code_servers key lets you configure how Dagster loads the code in your workspace.

When you aren't running your own gRPC server, Dagit and the Dagster Daemon load your code from a gRPC server running in a subprocess. By default, if your code takes more than 60 seconds to load, Dagster will assume that it is hanging and stop waiting for it to load. If you expect that your repository code will take longer than 60 seconds to load, you can set the local_startup_timeout key:

# Configures how long Dagster waits for repositories
# to load before timing out.
code_servers:
  local_startup_timeout: 120

Data retention#

The retention key lets you configure how long Dagster retains certain types of data that have diminishing value over time, like schedule/sensor tick data. If you want to clean up old ticks to minimize storage concerns and improve query performance, you can set retention policy using the retention config key:

For schedule and sensor ticks, you can specify the field purge_after_days, which takes either a mapping of tick types to integers, or an integer that applies to all tick types. This integer determines after how many days that ticks can be safely removed. A value of -1 indicates that ticks should be retained indefinitely.

# Configures how long Dagster keeps sensor / schedule tick data
retention:
  schedule:
    purge_after_days: 90 # sets retention policy for schedule ticks of all types
  sensor:
    purge_after_days:
      skipped: 7
      failure: 30
      success: -1 # keep success ticks indefinitely

By default, Dagster retains skipped sensor ticks for 7 days and retains all other ticks indefinitely.

Sensor evaluation#

The sensors key lets you configure how your sensors get evaluated. If you want your sensors to be evaluated asynchronously, you can set the use_threads attribute as well as a num_workers config setting.

sensors:
  use_threads: true
  num_workers: 8

Schedule evaluation#

Likewise, the schedules key lets you configure how your schedules get evaluated. If you want your schedules to be evaluated asynchronously, you can set the use_threads attribute as well as a num_workers config setting.

schedules:
  use_threads: true
  num_workers: 8

By default, Dagster evaluates sensors synchronously.