Fault-tolerant execution#

By default, if a Trino node lacks the resources to execute a task or otherwise fails during query execution, the query fails and must be run again manually. The longer the runtime of a query, the more likely it is to be susceptible to such failures.

Fault-tolerant execution is a mechanism in Trino that enables a cluster to mitigate query failures by retrying queries or their component tasks in the event of failure. With fault-tolerant execution enabled, intermediate exchange data is spooled and can be re-used by another worker in the event of a worker outage or other fault during query execution.

Note

Fault tolerance does not apply to broken queries or other user error. For example, Trino does not spend resources retrying a query that fails because its SQL cannot be parsed.

Note

Fault-tolerant execution is a public preview in Starburst Enterprise. Contact Starburst Support with questions or feedback.

Configuration#

Fault-tolerant execution is disabled by default. To enable the feature, set the retry-policy configuration property to either QUERY or TASK depending on the desired retry policy.

retry-policy=QUERY

Warning

Setting retry-policy disables write operations with connectors that do not support fault-tolerant execution of write operations, resulting in a “This connector does not support query retries” error message.

Support for fault-tolerant execution of SQL statements varies on a per-connector basis:

The following configuration properties control the behavior of fault-tolerant execution on a Trino cluster:

Fault-tolerant execution configuration properties#

Property name

Description

Default value

retry-policy

Configures what is retried in the event of failure, either QUERY to retry the whole query, or TASK to retry tasks individually if they fail. See retry policy for more information.

NONE

exchange.deduplication-buffer-size

Size of the coordinator’s in-memory buffer used by fault-tolerant execution to store output of query stages. If this buffer is filled during query execution, the query fails with a “Task descriptor storage capacity has been exceeded” error message unless an exchange manager is configured.

32MB

exchange.compression-enabled

Enable compression of spooling data. Setting to true is recommended when using an exchange manager.

false

Retry policy#

The retry-policy configuration property designates whether Trino retries entire queries or a query’s individual tasks in the event of failure.

QUERY#

A QUERY retry policy instructs Trino to automatically retry a query in the event of an error occuring on a worker node. A QUERY retry policy is recommended when the majority of the Trino cluster’s workload consists of many small queries, or if an exchange manager is not configured.

By default Trino does not implement fault tolerance for queries whose result set exceeds 32MB in size, such as SELECT statements that return a very large data set to the user. This limit can be increased by modifying the exchange.deduplication-buffer-size configuration property to be greater than the default value of 32MB, but this results in higher memory usage on the coordinator.

To enable fault-tolerant execution on queries with a larger result set, it is strongly recommended to configure an exchange manager that utilizes external storage for spooled data and therefore allows for storage of spilled data beyond the in-memory buffer size.

TASK#

A TASK retry policy instructs Trino to retry individual query tasks in the event of failure. This policy is recommended when executing large batch queries, as the cluster can more efficiently retry smaller tasks within the query rather than retry the whole query.

The following cluster configuration changes are recommended to improve fault-tolerant execution with a TASK retry policy:

  • Set the task.low-memory-killer.policy query management property to total-reservation-on-blocked-nodes, or queries may need to be manually killed if the cluster runs out of memory.

  • Set the query.low-memory-killer.delay query management property to 0s so the cluster immediately unblocks nodes that run out of memory.

  • Modify the query.remote-task.max-error-duration query management property to adjust how long Trino allows a remote task to try reconnecting before considering it lost and rescheduling.

Note

A TASK retry policy is best suited for large batch queries, but this policy can result in higher latency for short-running queries executed in high volume. As a best practice, it is recommended to run a dedicated cluster with a TASK retry policy for large batch queries, separate from another cluster that handles short queries.

Advanced configuration#

You can further configure fault-tolerant execution with the following configuration properties. The default values for these properties should work for most deployments, but you can change these values for testing or troubleshooting purposes.

Retry limits#

The following configuration properties control the thresholds at which queries/tasks are no longer retried in the event of repeated failures:

Fault tolerance retry limit configuration properties#

Property name

Description

Default value

Retry policy

query-retry-attempts

Maximum number of times Trino may attempt to retry a query before declaring the query as failed.

4

Only QUERY

task-retry-attempts-overall

Maximum number retries across all tasks within a given query before declaring the query as failed.

null (no limit)

Only TASK

task-retry-attempts-per-task

Maximum number of times Trino may attempt to retry a single task before declaring the query as failed.

4

Only TASK

retry-initial-delay

Minimum time that a failed query or task must wait before it is retried. May be overridden with the retry_initial_delay session property.

10s

QUERY and TASK

retry-max-delay

Maximum time that a failed query or task must wait before it is retried. Wait time is increased on each subsequent failure. May be overridden with the retry_max_delay session property.

1m

QUERY and TASK

retry-delay-scale-factor

Factor by which retry delay is increased on each query or task failure. May be overridden with the retry_delay_scale_factor session property.

2.0

QUERY and TASK

Task sizing#

With a TASK retry policy, it is important to manage the amount of data processed in each task. If tasks are too small, the management of task coordination can take more processing time and resources than executing the task itself. If tasks are too large, then a single task may require more resources than are available on any one node and therefore prevent the query from completing.

Trino supports limited automatic task sizing. If issues are occurring during fault-tolerant task execution, you can configure the following configuration properties to manually control task sizing. These configuration properties only apply to a TASK retry policy.

Task sizing configuration properties#

Property name

Description

Default value

fault-tolerant-execution-target-task-input-size

Target size in bytes of all task inputs for a single fault-tolerant task. Applies to tasks that read input from spooled data written by other tasks.

May be overridden for the current session with the fault_tolerant_execution_target_task_input_size session property.

4GB

fault-tolerant-execution-target-task-split-count

Target number of standard splits processed by a single task that reads data from source tables. Value is interpreted with split weight taken into account. If the weight of splits produced by a catalog denotes that they are lighter or heavier than “standard” split, then the number of splits processed by single task is adjusted accordingly.

May be overridden for the current session with the fault_tolerant_execution_target_task_split_count session property.

64

fault-tolerant-execution-min-task-split-count

Minimum number of splits processed by a single task. This value is not split weight-adjusted and serves as protection against situations where catalogs report an incorrect split weight.

May be overridden for the current session with the fault_tolerant_execution_min_task_split_count session property.

16

fault-tolerant-execution-max-task-split-count

Maximum number of splits processed by a single task. This value is not split weight-adjusted and serves as protection against situations where catalogs report an incorrect split weight.

May be overridden for the current session with the fault_tolerant_execution_max_task_split_count session property.

256

Node allocation#

With a TASK retry policy, nodes are allocated to tasks based on available memory and estimated memory usage. If task failure occurs due to exceeding available memory on a node, the task is restarted with a request to allocate the full node for its execution.

The initial task memory-requirements estimation is static and configured with the fault-tolerant-task-memory configuration property. This property only applies to a TASK retry policy.

Node allocation configuration properties#

Property name

Description

Default value

fault-tolerant-execution-task-memory

Initial task memory estimation used for bin-packing when allocating nodes for tasks. May be overridden for the current session with the fault_tolerant_execution_task_memory session property.

5GB

Other tuning#

The following additional configuration property can be used to manage fault-tolerant execution:

Other fault-tolerant execution configuration properties#

Property name

Description

Default value

Retry policy

fault-tolerant-execution-task-descriptor-storage-max-memory

Maximum amount of memory to be used to store task descriptors for fault tolerant queries on coordinator. Extra memory is needed to be able to reschedule tasks in case of a failure.

(JVM heap size * 0.15)

Only TASK

fault-tolerant-execution-partition-count

Number of partitions to use for distributed joins and aggregations, similar in function to the query.hash-partition-count query management property. It is not recommended to increase this property value above the default of 50, which may result in instability and poor performance. May be overridden for the current session with the fault_tolerant_execution_partition_count session property.

50

Only TASK

max-tasks-waiting-for-node-per-stage

Allow for up to configured number of tasks to wait for node allocation per stage, before pausing scheduling for other tasks from this stage.

5

Only TASK

Exchange manager#

Exchange spooling is responsible for storing and managing spooled data for fault-tolerant execution. You can configure a filesystem-based exchange manager that stores spooled data in a specified location, such as an S3-compatible storage system, Google Cloud Storage (GCS), or a local filesystem.

Configuration#

To configure an exchange manager, create a new etc/exchange-manager.properties configuration file on the coordinator and all worker nodes. In this file, set the exchange-manager.name configuration propertry to filesystem, and additional configuration properties as needed for your storage solution.

The following table lists the available configuration properties for exchange-manager.properties, their default values, and which filesystem(s) the property may be configured for:

Exchange manager configuration properties#

Property name

Description

Default value

Supported filesystem

exchange.base-directories

Comma-separated list of URI locations that the exchange manager uses to store spooling data. Only supports S3 and local filesystems.

Any

exchange.encryption-enabled

Enable encrypting of spooling data.

true

Any

exchange.sink-buffer-pool-min-size

The minimum buffer pool size for an exchange sink. The larger the buffer pool size, the larger the write parallelism and memory usage.

10

Any

exchange.sink-buffers-per-partition

The number of buffers per partition in the buffer pool. The larger the buffer pool size, the larger the write parallelism and memory usage.

2

Any

exchange.sink-max-file-size

Max size of files written by exchange sinks.

1GB

Any

exchange.source-concurrent-reader

Number of concurrent readers to read from spooling storage. The larger the number of concurrent readers, the larger the read parallelism and memory usage.

4

Any

exchange.s3.aws-access-key

AWS access key to use. Required for a connection to AWS S3 and GCS, can be ignored for other S3 storage systems.

AWS S3, GCS

exchange.s3.aws-secret-key

AWS secret key to use. Required for a connection to AWS S3 and GCS, can be ignored for other S3 storage systems.

AWS S3, GCS

exchange.s3.iam-role

IAM role to assume.

AWS S3, GCS

exchange.s3.external-id

External ID for the IAM role trust policy.

AWS S3, GCS

exchange.s3.region

Region of the S3 bucket.

AWS S3, GCS

exchange.s3.endpoint

S3 storage endpoint server if using an S3-compatible storage system that is not AWS. If using AWS S3, this can be ignored. If using GCS, set it to https://storage.googleapis.com.

Any S3-compatible storage

exchange.s3.max-error-retries

Maximum number of times the exchange manager’s S3 client should retry a request.

10

Any S3-compatible storage

exchange.s3.upload.part-size

Part size for S3 multi-part upload.

5MB

Any S3-compatible storage

exchange.gcs.json-key-file-path

Path to the JSON file that contains your Google Cloud Platform service account key. Not to be set together with exchange.gcs.json-key

GCS

exchange.gcs.json-key

Your Google Cloud Platform service account key in JSON format. Not to be set together with exchange.gcs.json-key-file-path

GCS

exchange.azure.connection-string

Connection string used to access the spooling container.

Azure Blob Storage

exchange.azure.block-size

Block size for Azure block blob parallel upload.

4MB

Azure Blob Storage

exchange.azure.max-error-retries

Maximum number of times the exchange manager’s Azure client should retry a request.

10

Azure Blob Storage

It is recommended to set the exchange.compression-enabled property to true in the cluster’s config.properties file, to reduce the exchange manager’s overall I/O load. It is also recommended to configure a bucket lifecycle rule to automatically expire abandoned objects in the event of a node crash.

AWS S3#

The following example exchange-manager.properties configuration specifies an AWS S3 bucket as the spooling storage destination. Note that the destination does not have to be in AWS, but can be any S3-compatible storage system.

exchange-manager.name=filesystem
exchange.base-directories=s3://exchange-spooling-bucket
exchange.s3.region=us-west-1
exchange.s3.aws-access-key=example-access-key
exchange.s3.aws-secret-key=example-secret-key

You can configure multiple S3 buckets for the exchange manager to distribute spooled data across buckets, reducing the I/O load on any one bucket. If a query fails with the error message “software.amazon.awssdk.services.s3.model.S3Exception: Please reduce your request rate”, this indicates that the workload is I/O intensive, and you should specify multiple S3 buckets in exchange.base-directories to balance the load:

exchange.base-directories=s3://exchange-spooling-bucket-1,s3://exchange-spooling-bucket-2

Azure Blob Storage#

The following example exchange-manager.properties configuration specifies an Azure Blob Storage container as the spooling storage destination.

exchange-manager.name=filesystem
exchange.base-directories=abfs://container_name@account_name.dfs.core.windows.net
exchange.azure.connection-string=connection-string

Google Cloud Storage#

To enable exchange spooling on GCS in Trino, change the request endpoint to the https://storage.googleapis.com Google storage URI, and configure your AWS access/secret keys to use the GCS HMAC keys. If you deploy Trino on GCP, you must either create a service account with access to your spooling bucket or configure the key path to your GCS credential file.

For more information on GCS’s S3 compatibility, refer to the Google Cloud documentation on S3 migration.

The following example exchange-manager.properties configuration specifies a GCS bucket as the spooling storage destination.

exchange-manager.name=filesystem
exchange.base-directories=gs://exchange-spooling-bucket
exchange.s3.region=us-west-1
exchange.s3.aws-access-key=example-access-key
exchange.s3.aws-secret-key=example-secret-key
exchange.s3.endpoint=https://storage.googleapis.com
exchange.gcs.json-key-file-path=/path/to/gcs_keyfile.json

Local filesystem storage#

The following example exchange-manager.properties configuration specifies a local directory, /tmp/trino-exchange-manager, as the spooling storage destination.

Note

It is only recommended to use a local filesystem for exchange in standalone, non-production clusters. A local directory can only be used for exchange in a distributed cluster if the exchange directory is shared and accessible from all worker nodes.

exchange-manager.name=filesystem
exchange.base-directories=/tmp/trino-exchange-manager