Streaming Parallelism

Streaming parallelism is configured by two kinds of session parameters:

• streaming_parallelism
• streaming_parallelism_for_<type>

Supported <type> values in this document:

• table
• materialized_view
• index
• source
• sink

Value Semantics

The unified parallelism value accepts the following forms:

• 0
• adaptive
• default
• <n>
• bounded(<n>)
• ratio(<r>)

The meaning is:

• 0 and adaptive mean adaptive scheduling with the full worker set.
• <n> means fixed parallelism.
• bounded(<n>) means adaptive scheduling with an upper bound.
• ratio(<r>) means adaptive scheduling with a ratio of the available worker parallelism.

The stored default value for all unified parameters is default.

The effective defaults preserve the legacy behavior:

• streaming_parallelism = default, which resolves to bounded(64)
• streaming_parallelism_for_table = default, which resolves to bounded(4) only when streaming_parallelism is also default
• streaming_parallelism_for_source = default, which resolves to bounded(4) only when streaming_parallelism is also default
• all other streaming_parallelism_for_<type> values default to default and follow streaming_parallelism

For these parameters, SET ... = DEFAULT and SET ... = 'default' both restore the stored default value, so SHOW output remains round-trippable.

Resolution Rules

Resolution happens in a single step:

1. Resolve streaming_parallelism.
   • default resolves to the legacy global default bounded(64).
2. Resolve streaming_parallelism_for_<type>.
   • For table and source, default resolves to the legacy type-specific default bounded(4) only when streaming_parallelism is still default; otherwise it follows the resolved global value.
   • For other job types, default falls back to the resolved global value.
3. Convert the resolved value into:
   • a fixed parallelism, or
   • an adaptive strategy stored in the streaming job metadata.

This means the frontend always writes the final adaptive strategy into the job context, and meta no longer depends on a separate system-wide adaptive strategy parameter.

Examples

SET streaming_parallelism = adaptive;
SET streaming_parallelism = 8;
SET streaming_parallelism = 'bounded(16)';
SET streaming_parallelism = 'ratio(0.5)';

SET streaming_parallelism = 'ratio(0.5)';
SET streaming_parallelism_for_materialized_view = 'bounded(4)';
SET streaming_parallelism_for_materialized_view = DEFAULT;

Migration

The deprecated parameters fall into two groups:

1. adaptive_parallelism_strategy

   This parameter existed in official releases and was the old cluster-wide knob for adaptive scheduling. Users could previously set it through ALTER SYSTEM or config files to choose how adaptive should expand by default, for example AUTO, BOUNDED(n), or RATIO(r).

   This is no longer supported as a separate user-facing setting. After the migration, users must express the final policy directly with streaming_parallelism and streaming_parallelism_for_<type>, such as adaptive, bounded(64), or ratio(0.5). There is no longer a separate "adaptive strategy default" to configure.

2. streaming_parallelism_strategy and streaming_parallelism_strategy_for_<type>

   These parameters only existed in Nightly builds between v2.8.0 and v3.0.0, so this part of the migration is not a breaking change for official releases. It still matters for Nightly users because their stored values are migrated to the unified parameters.

   These parameters were also separate strategy-only knobs: users could combine streaming_parallelism = adaptive with a matching streaming_parallelism_strategy = ..., or do the same per job type. This split representation is no longer supported. After the migration, users set only the unified streaming_parallelism* parameters, and each value already carries its own strategy.

On startup, meta derives the final streaming_parallelism and streaming_parallelism_for_<type> values once from the legacy parameters, persists the new values, and drops the deprecated entries. For adaptive_parallelism_strategy, if the legacy system parameter is missing, the migration interprets it using the released legacy default AUTO. Existing jobs without a stored job-level strategy are materialized as AUTO to preserve their released behavior, while untouched sessions on the new runtime still resolve the unified defaults (streaming_parallelism = default to bounded(64), table/source default to bounded(4)). New clusters only expose the unified parameters.

In other words, an upgraded user may still observe the same effective parallelism for existing jobs or migrated session defaults, but they can no longer keep adjusting that behavior through the old *_strategy parameters. Any further changes must use the unified streaming_parallelism* parameters.