Skip to main content

Datasets

A Spicepod can contain one or more datasets referenced by relative path or defined inline.

Inline example:

spicepod.yaml

datasets:
  - from: spice.ai/spiceai/quickstart/datasets/taxi_trips
    name: taxi_trips
    acceleration:
      enabled: true
      mode: memory # / file
      engine: arrow # / cayenne / duckdb / sqlite / postgres
      refresh_check_interval: 1h
      refresh_mode: full / append # update / incremental

spicepod.yaml

datasets:
  - from: databricks:spiceai.datasets.specific_table
    name: uniswap_eth_usd
    params:
      environment: prod
    acceleration:
      enabled: true
      mode: memory # / file
      engine: arrow # / duckdb
      refresh_check_interval: 1h
      refresh_mode: full / append # update / incremental

Relative path example:

spicepod.yaml

datasets:
  - ref: datasets/taxi_trips

datasets/taxi_trips/dataset.yaml

from: spice.ai/spiceai/quickstart/datasets/taxi_trips
name: taxi_trips
type: overwrite
acceleration:
  enabled: true
  refresh: 1h

from​

The from field is a string that represents the Uniform Resource Identifier (URI) for the dataset. This URI is composed of two parts: a prefix indicating the Data Connector to use to connect to the dataset, a delimiter, and the path to the dataset within the source.

The syntax for the from field is as follows:

from: <data_connector>:<path>
# OR
from: <data_connector>/<path>
# OR
from: <data_connector>://<path>

Where:

  • <data_connector>: The Data Connector to use to connect to the dataset

    Currently supported data connectors:

    If the Data Connector is not explicitly specified, it defaults to spiceai.

  • <delimiter>: The delimiter between the Data Connector and the path. Currently supported delimiters are :, /, and ://. Some connectors place additional restrictions on the allowed delimiters to better conform to the expected syntax of the underlying data source, i.e. s3:// is the only supported delimiter for the s3 connector.

  • <path>: The path to the dataset within the source.

ref​

An alternative to adding the dataset definition inline in the spicepod.yaml file. ref can be use to point to a directory with a dataset defined in a dataset.yaml file. For example, a dataset configured in a dataset.yaml in the "datasets/sample" directory can be referenced with the following:

dataset.yaml

from: spice.ai/spiceai/quickstart/datasets/taxi_trips
name: taxi_trips
type: overwrite
acceleration:
  enabled: true
  refresh: 1h

ref used in spicepod.yaml

version: v1
kind: Spicepod
name: duckdb
datasets:
  - ref: datasets/sample

name​

The name of the dataset. Used to reference the dataset in the pod manifest, as well as in external data sources. The name cannot be a reserved keyword.

description​

The description of the dataset. Used as part of the Semantic Data Model.

access​

Optional. Specifies the access level for the dataset. Supported values are:

  • read (default): Read-only access.
  • read_write: Enables both read and write operations. Only supported for write-capable connectors.

To enable write operations, configure your dataset with read_write access:

datasets:
  - from: glue:my_catalog.my_schema.my_table
    name: my_table
    access: read_write
    params:
      # ... connector-specific parameters

time_column​

Optional. The name of the column that represents the temporal (time) ordering of the dataset.

Required to enable a retention policy on the dataset.

time_format​

Optional. The format of the time_column. The following values are supported:

  • timestamp - Default. Timestamp without a timezone. E.g. 2016-06-22 19:10:25 with data type timestamp.
  • timestamptz - Timestamp with a timezone. E.g. 2016-06-22 19:10:25-07 with data type timestamptz.
  • unix_seconds - Unix timestamp in seconds. E.g. 1718756687.
  • unix_millis - Unix timestamp in milliseconds. E.g. 1718756687000.
  • ISO8601 - ISO 8601 format.
  • date - Date in YYYY-MM-DD format. E.g. 2024-01-01.

Spice emits a warning if the time_column from the data source is incompatible with the time_format config.

Limitations
  • String-based columns are assumed to be ISO8601 format.

time_partition_column​

(Optional) Specify the column that represents the physical partitioning of the dataset when using append-based acceleration. When the defined time_column is a fine-grained timestamp and the dataset is physically partitioned by a coarser granularity (for example, by date), setting time_partition_column to the partition column (e.g. date_col) improves partition pruning, excludes irrelevant partitions during refreshes, and optimizes scan efficiency.

time_partition_format​

(Optional) Define the format of the time_partition_column. For instance, if the physical partitions follow a date format (YYYY-MM-DD), set this value to date. The same format options as time_format are supported for time_partition_column.

unsupported_type_action​

Optional. Specifies the action to take when a data type that is not supported by the data connector is encountered.

The following values are supported:

  • error - Default. Return an error when an unsupported data type is encountered.
  • warn - Log a warning and ignore the column containing the unsupported data type.
  • ignore - Log nothing and ignore the column containing the unsupported data type.
  • string - Attempt to convert the unsupported data type to a string. Currently only supports converting the PostgreSQL JSONB type.
Limitations

Not all connectors support specifying an unsupported_type_action. When specified on a connector that does not support the option, the connector will fail to register. The following connectors support unsupported_type_action:

ready_state​

Supports one of two values:

  • on_registration: Mark the dataset as ready immediately, and queries on this table will fall back to the underlying source directly until the initial acceleration is complete
  • on_load: Mark the dataset as ready only after the initial acceleration. Queries against the dataset will return an error before the load has been completed.
datasets:
  - from: s3://my_bucket/my_dataset/
    name: my_dataset
    ready_state: on_registration # or on_load
    params: ...
    acceleration:
      enabled: true

check_availability​

Spice monitors the availability of non-accelerated datasets and emits metrics if a dataset becomes unavailable. Note that this monitoring process may trigger the startup of compute resources (for example, Databricks or Snowflake), potentially incurring additional costs. To disable availability monitoring, configure the check_availability parameter to disabled.

  • auto: Automatically check the availability monitor of the dataset. This is the default value. Accelerated datasets are not monitored.
  • disabled: Disable the availability monitor for the dataset.
datasets:
  - from: databricks:catalog.schema.table
    name: my_dataset
    check_availability: disabled
    params: ...

The monitoring works by executing a query that selects one row and all columns from the dataset. i.e.:

SELECT
  "p_partkey",
  "p_name",
  "p_mfgr",
  "p_brand",
  "p_type",
  "p_size",
  "p_container",
  "p_retailprice",
  "p_comment"
FROM
  spiceai_sandbox.tpch.part
LIMIT 1

If the monitoring query fails a warning is emitted in the logs, an error is propagated to the task_history table and the dataset_unavailable_time_ms metric is incremented for the failing dataset.

acceleration​

Optional. Accelerate queries to the dataset by caching data locally.

acceleration.enabled​

Enable or disable acceleration, defaults to true.

acceleration.engine​

The acceleration engine to use, defaults to arrow. The following engines are supported:

  • arrow - Accelerated in-memory backed by Apache Arrow DataTables.
  • cayenne - Accelerated by the Cayenne (Vortex) engine (Alpha, v1.9.0-rc.1+).
  • duckdb - Accelerated by an embedded DuckDB database.
  • postgres - Accelerated by a Postgres database.
  • sqlite - Accelerated by an embedded SQLite database.

acceleration.mode​

Optional. The mode of acceleration. The following values are supported:

  • memory - Store acceleration data in-memory. Not supported for the cayenne engine.
  • file - Store acceleration data in a file. Supported for cayenne, duckdb and sqlite acceleration engines.

acceleration.snapshots​

Optional. Controls how this dataset participates in managed acceleration snapshots. Requires the Spicepod to configure the top-level snapshots block, the acceleration engine to be duckdb or sqlite, and mode: file with a dataset-specific file path (for example acceleration.params.duckdb_file: /nvme/my_dataset.db).

Supported values:

  • enabled – Download the newest snapshot on startup when the acceleration file is missing and write a fresh snapshot after each refresh.
  • bootstrap_only – Download snapshots on startup but never write new ones.
  • create_only – Write snapshots after refreshes but never download them on startup.
  • disabled (default) – Do not use snapshots for this dataset.

Snapshots are written beneath the configured snapshot location using Hive-style partitioning (month=YYYY-MM/day=YYYY-MM-DD/dataset=<dataset>). For more background, see Acceleration snapshots.

acceleration.refresh_mode​

Optional. How to refresh the dataset. The following values are supported:

  • full - Refresh the entire dataset.
  • append - Append new data to the dataset. When time_column is specified, new records are fetched from the latest timestamp in the accelerated data at the acceleration.refresh_check_interval.
  • changes - Apply change data capture (CDC) events to incrementally update the dataset.
  • caching - Cache data based on request metadata (HTTP requests). Uses row-level replacement based on cache keys. See Caching Mode for details.

acceleration.refresh_check_interval​

Optional. How often data should be refreshed. For append datasets without a specific time_column, this config is not used. If not defined, the accelerator will not refresh after it initially loads data. Cannot be specified in conjunction with a refresh_cron.

See Duration

acceleration.refresh_cron​

Optional. Specifies a cron schedule which controls how often data is refreshed. For append datasets without a specific time_column, this config is not used. If not defined, the accelerator will not refresh after it initially loads data.

See the cron schedule reference.

acceleration.params.caching_ttl​

Optional. The time-to-live (TTL) for cached data before it is considered stale. Only applicable when refresh_mode: caching. Defaults to 30s.

When cached data exceeds this age (measured from the fetched_at timestamp), it becomes stale. If caching_stale_while_revalidate_ttl is also configured, stale data is immediately served to queries (no delay) while a background refresh is triggered to update the cache, implementing the Stale-While-Revalidate (SWR) pattern. If caching_stale_while_revalidate_ttl is not set, queries wait for fresh data once the TTL expires.

Example:

datasets:
  - from: https://api.tvmaze.com
    name: tv_shows
    acceleration:
      enabled: true
      refresh_mode: caching
      engine: duckdb
      mode: file # Persist cache to disk
      params:
        caching_ttl: 15s # Cache data is fresh for 15 seconds
      refresh_check_interval: 30s # Periodic background refresh

See Caching Mode for detailed TTL configuration and behavior.

See Duration

acceleration.params.caching_stale_while_revalidate_ttl​

Optional. The duration after caching_ttl expires during which stale data is served while refreshing in the background. Only applicable when refresh_mode: caching. Defaults to none (stale data is not served).

When caching_ttl expires and data becomes stale, this parameter controls how long stale data continues to be served immediately while a background refresh occurs. After the combined caching_ttl + caching_stale_while_revalidate_ttl period, queries wait for fresh data instead of returning stale results.

If omitted, cached data becomes "rotten" immediately after caching_ttl expires, and queries will wait for fresh data rather than returning stale results.

Example:

datasets:
  - from: https://api.tvmaze.com
    name: tv_shows
    acceleration:
      enabled: true
      refresh_mode: caching
      engine: duckdb
      mode: file
      params:
        caching_ttl: 15s # Cache data is fresh for 15 seconds
        caching_stale_while_revalidate_ttl: 30s # Serve stale data for 30 seconds while refreshing
      refresh_check_interval: 60s

See Caching Mode for detailed TTL configuration and behavior.

See Duration

acceleration.params.caching_stale_if_error​

Optional. Controls whether expired cached data is served when the upstream data source returns an error. Only applicable when refresh_mode: caching. Defaults to disabled.

When set to enabled, queries return expired cached data instead of failing if the upstream source returns an error during a refresh attempt. This provides fault tolerance for APIs with intermittent availability or rate limits.

Valid values:

  • enabled - Serve expired cached data when upstream errors occur
  • disabled (default) - Propagate upstream errors to queries

Example:

datasets:
  - from: https://api.tvmaze.com
    name: tv_shows
    acceleration:
      enabled: true
      refresh_mode: caching
      engine: duckdb
      mode: file
      params:
        caching_ttl: 15s
        caching_stale_while_revalidate_ttl: 30s
        caching_stale_if_error: enabled # Serve stale data on upstream errors
      refresh_check_interval: 60s

See Caching Mode for detailed behavior.

acceleration.refresh_sql​

Optional. Filters the data fetched from the source to be stored in the accelerator engine. Only supported for full refresh_mode datasets.

Must be of the form SELECT * FROM {name} WHERE {refresh_filter}. {name} is the dataset name declared above, {refresh_filter} is any SQL expression that can be used to filter the data, i.e. WHERE city = 'Seattle' to reduce the working set of data that is accelerated within Spice from the data source.

Limitations
  • The refresh SQL only supports filtering data from the current dataset - joining across other datasets is not supported.
  • Queries for data that have been filtered out will not fall back to querying against the federated table.

acceleration.refresh_data_window​

Optional. A duration to filter dataset refresh source queries to recent data (duration into past from now). Requires time_column and time_format to also be configured. Only supported for full refresh mode datasets.

For example, refresh_data_window: 24h will include only records with a timestamp within the last 24 hours.

See Duration

acceleration.refresh_append_overlap​

Optional. A duration to specify how far back to include records based on the most recent timestamp found in the accelerated data. Requires time_column to also be configured. Only supported for append refresh mode datasets.

This setting can help mitigate missing data issues caused by late arriving data.

Example: If the latest timestamp in the accelerated data table is 2020-01-01T02:00:00Z, setting refresh_append_overlap: 1h will include records starting from 2020-01-01T01:00:00Z.

See Duration

acceleration.refresh_retry_enabled​

Optional. Specifies whether an accelerated dataset should retry data refresh in the event of transient errors. The default setting is true.

Retries utilize a Fibonacci backoff strategy. To disable refresh retries, set refresh_retry_enabled: false.

acceleration.refresh_retry_max_attempts​

Optional. Defines the maximum number of retry attempts when refresh retries are enabled. The default is undefined, allowing for unlimited attempts.

acceleration.refresh_on_startup​

Optional. Controls the refresh behavior of an accelerated dataset across restarts. Defaults to auto.

Supported Values​

  • auto (Default) – Maintains refresh state across restarts:
    • With refresh_check_interval: Schedules next refresh based on last successful refresh time, triggering immediately if interval has already elapsed
    • Without refresh_check_interval: No refresh (on-demand only)
  • always – Forces a dataset refresh on every startup, regardless of the existing acceleration state.

Setting refresh_on_startup: always ensures that accelerated data is always refreshed to match the source when the service restarts. This is useful in development environments or when data consistency is critical after deployment.

acceleration.params​

Optional. Parameters to pass to the acceleration engine. The parameters are specific to the acceleration engine used.

acceleration.engine_secret​

Optional. The secret store key to use the acceleration engine connection credential. For supported data connectors, use spice login to store the secret.

acceleration.retention_check_enabled​

Optional. Enable or disable retention policy check, defaults to false.

acceleration.retention_period​

Optional. The retention period for the dataset. Combine with time_column and time_format to determine if the data should be retained or not.

retention_period or retention_sql must be specified when acceleration.retention_check_enabled is true. When both retention_period and retention_sql are configured, both retention policies will be applied during each retention check.

See Duration

acceleration.retention_sql​

Optional. Custom SQL statement to define data retention logic. Takes the form of a DELETE FROM <table> WHERE <predicates> statement.

This parameter is useful for scenarios like soft-deleting rows in append-only datasets or removing data based on complex business logic that goes beyond simple time-based retention.

retention_period or retention_sql must be specified when acceleration.retention_check_enabled is true. When both retention_period and retention_sql are configured, both retention policies will be applied during each retention check.

acceleration.retention_check_interval​

Optional. How often the retention policy should be checked.

Required when acceleration.retention_check_enabled is true.

See Duration

acceleration.refresh_jitter_enabled​

Optional. Enable or disable refresh jitter, defaults to false. The refresh jitter adds/substracts a randomized time period from the refresh_check_interval.

acceleration.refresh_jitter_max​

Optional. The maximum amount of jitter to add to the refresh interval. The jitter is a random value between 0 and refresh_jitter_max. Defaults to 10% of refresh_check_interval.

metrics​

Optional. Enable component-specific metrics for the dataset. Each component can expose its own set of metrics that can be enabled selectively to monitor specific aspects of its operation.

Component metrics are disabled by default and can be enabled by adding a metrics section to the dataset configuration. Each metric can be enabled individually by specifying its name in the metrics list.

Example Configuration​

datasets:
  - from: mysql:my_table
    name: my_dataset
    metrics:
      - name: connection_count
        enabled: true
      - name: connections_in_pool
        enabled: true
      - name: active_wait_requests
        enabled: true
    params:
      mysql_host: localhost
      mysql_tcp_port: 3306
      mysql_user: root
      mysql_pass: ${secrets:MYSQL_PASS}

For detailed information about metrics available for specific components, see the component metrics documentation.

acceleration.indexes​

Optional. Specify which indexes should be applied to the locally accelerated table. Not supported for in-memory Arrow acceleration engine.

The indexes field is a map where the key is the column reference and the value is the index type.

A column reference can be a single column name or a multicolumn key. The column reference must be enclosed in parentheses if it is a multicolumn key.

See Indexes

datasets:
  - from: spice.ai/eth.recent_blocks
    name: eth.recent_blocks
    acceleration:
      enabled: true
      engine: sqlite
      indexes:
        number: enabled # Index the `number` column
        '(hash, timestamp)': unique # Add a unique index with a multicolumn key comprised of the `hash` and `timestamp` columns

acceleration.primary_key​

Optional. Specify the primary key constraint on the locally accelerated table. Not supported for in-memory Arrow acceleration engine.

The primary_key field is a string that represents the column reference that should be used as the primary key. The column reference can be a single column name or a multicolumn key. The column reference must be enclosed in parentheses if it is a multicolumn key.

See Constraints

datasets:
  - from: spice.ai/eth.recent_blocks
    name: eth.recent_blocks
    acceleration:
      enabled: true
      engine: sqlite
      primary_key: hash # Define a primary key on the `hash` column

acceleration.on_conflict​

Optional. Specify what should happen when a constraint is violated. Not supported for in-memory Arrow acceleration engine.

The on_conflict field is a map where the key is the column reference and the value is the conflict resolution strategy.

A column reference can be a single column name or a multicolumn key. The column reference must be enclosed in parentheses if it is a multicolumn key.

Only a single on_conflict target can be specified, unless all on_conflict targets are specified with drop.

The possible conflict resolution strategies are:

  • upsert - Upsert the incoming data when the primary key constraint is violated.
  • upsert_dedup - Same as upsert, but also deduplicates the data if there are duplicate rows that trigger a violation constraint within a single update. See Advanced upsert behavior.
  • upsert_dedup_by_row_id - Same as upsert, but resolves any violations by arbitrarily choosing the row with the highest row id. See Advanced upsert behavior.
  • drop - Drop the data when the primary key constraint is violated.

See Constraints

acceleration.snapshots_trigger_threshold​

Optional. Specify how frequently snapshots are created during streaming operations. Only applicable when using refresh_mode: changes and refresh_mode: append without time_column.

The snapshots_trigger_threshold field is an integer that determines after how many batch updates a snapshot should be created. For example, a value of 5 means a snapshot will be created every 5 batch updates.

datasets:
  - from: dynamodb:my_table
    name: orders_stream
    acceleration:
      enabled: true
      refresh_mode: changes
      on_conflict:
        (id, version): upsert
      params:
        snapshots_trigger_threshold: 5 # Create snapshot every 5 batch updates

See Snapshots

datasets:
  - from: spice.ai/eth.recent_blocks
    name: eth.recent_blocks
    acceleration:
      enabled: true
      engine: sqlite
      primary_key: hash
      indexes:
        '(number, timestamp)': unique
      on_conflict:
        # Upsert the incoming data when the primary key constraint on "hash" is violated,
        # alternatively "drop" can be used instead of "upsert" to drop the data update.
        hash: upsert

columns​

Optional. Define metadata, semantic details and features (e.g. embeddings, or table indexes) for specific columns in the dataset.

datasets:
  - from: file:sales_data.parquet
    name: sales
    columns:
      - name: address_line1
        description: The first line of the address.
        embeddings:
          - from: hf_minilm
            row_id: order_number
            chunking:
              enabled: true
              target_chunk_size: 256
              overlap_size: 32
        full_text_search:
          enabled: true

columns[*].name​

The name of the column in the table schema.

columns[*].description​

Optional. A description of the column's contents and purpose. Used as part of the Semantic Data Model.

columns[*].embeddings​

Optional. Create vector embeddings for this column.

columns[*].embeddings[*].from​

The embedding model to use, specify the component name.

columns[*].embeddings[*].row_id​

Optional. For datasets without a primary key, used to explicitly specify column(s) that uniquely identify a row.

Specifying a row_id enables unique identifier lookups for datasets from external systems that may not have a primary key.

columns[*].embeddings[*].chunking​

Optional. The configuration to enable and define the chunking strategy for the embedding column.

columns:
  - name: description
    embeddings:
      - from: hf_minilm
        chunking:
          enabled: true
          target_chunk_size: 512
          overlap_size: 128
          trim_whitespace: false

See embeddings[*].chunking for details.

columns[*].embeddings[*].vector_size​

Optional. Specifies the size (number of dimensions) of the embedding vector for use in federated queries to databases that do not support arrays with fixed lengths.

columns:
  - name: review_body
    embeddings:
      - from: embed-static-retrieval
        vector_size: 1024

columns[*].full_text_search​

columns[*].full_text_search.enabled​

Optional. Enable or disable full text search support for specific column in the dataset. Default false.

columns[*].full_text_search.row_id​

Optional. For datasets without a primary key, used to explicitly specify column(s) that uniquely identify a row.

Specifying a row_id enables unique identifier lookups for datasets from external systems that may not have a primary key.

columns[*].metadata​

Optional. Specific metadata associated to the column.

columns[*].metadata.vectors​

Optional. If provided, a vector engine (see below) should store this column for a particular use, determined by the value, which is one of:

  • non-filterable: Store the column in the vector engine.
  • filterable: Store the column in the vector engine, and ensure the engine can filter on the column (if possible in the engine).

Only applicable if vectors.enabled is both defined and true.

embeddings​

Optional. Create vector embeddings for specific columns of the dataset.

datasets:
  - from: spice.ai/eth.recent_blocks
    name: eth.recent_blocks
    embeddings:
      - column: extra_data
        use: hf_minilm

embeddings[*].column​

The column name to create an embedding for.

embeddings[*].use​

The embedding model to use, specific the component name embeddings[*].name.

embeddings[*].column_pk​

Optional. For datasets without a primary key, explicitly specify column(s) that uniquely identify a row.

embeddings[*].chunking​

Optional. The configuration to enable and define the chunking strategy for the embedding column.

datasets:
  - from: spice.ai/eth.recent_blocks
    name: eth.recent_blocks
    embeddings:
      - column: extra_data
        use: hf_minilm
        chunking:
          enabled: true
          target_chunk_size: 512
          overlap_size: 128
          trim_whitespace: false

embeddings[*].chunking.enabled​

Optional. Enable or disable chunking for the embedding column. Defaults to false.

embeddings[*].chunking.target_chunk_size​

The desired size of each chunk, in tokens.

If the desired chunk size is larger than the maximum size of the embedding model, the maximum size will be used.

embeddings[*].chunking.overlap_size​

Optional. The number of tokens to overlap between chunks. Defaults to 0.

embeddings[*].chunking.trim_whitespace​

Optional. If enabled, the content of each chunk will be trimmed to remove leading and trailing whitespace. Defaults to true.

metadata​

Optional. Additional key-value metadata for the dataset. Used as part of the Semantic Data Model.

datasets:
  - from: spice.ai/eth.recent_blocks
    name: eth.recent_blocks
    metadata:
      instructions: The last 128 blocks.

vectors​

vectors.enabled​

Enable or disable vector storage, defaults to true.

vectors.engine​

The vector engine to use. The following engines are supported:

vectors.params​

Optional. Parameters to pass to the vector engine. The parameters are specific to the vector engine used.