diff --git a/mkdocs/docs/configuration.md b/mkdocs/docs/configuration.md index 06eaac1bed..e076afdb93 100644 --- a/mkdocs/docs/configuration.md +++ b/mkdocs/docs/configuration.md @@ -54,15 +54,18 @@ Iceberg tables support table properties to configure table behavior. ### Write options -| Key | Options | Default | Description | -| -------------------------------------- | --------------------------------- | ------- | ------------------------------------------------------------------------------------------- | -| `write.parquet.compression-codec` | `{uncompressed,zstd,gzip,snappy}` | zstd | Sets the Parquet compression coddec. | -| `write.parquet.compression-level` | Integer | null | Parquet compression level for the codec. If not set, it is up to PyIceberg | -| `write.parquet.row-group-limit` | Number of rows | 1048576 | The upper bound of the number of entries within a single row group | -| `write.parquet.page-size-bytes` | Size in bytes | 1MB | Set a target threshold for the approximate encoded size of data pages within a column chunk | -| `write.parquet.page-row-limit` | Number of rows | 20000 | Set a target threshold for the maximum number of rows within a column chunk | -| `write.parquet.dict-size-bytes` | Size in bytes | 2MB | Set the dictionary page size limit per row group | -| `write.metadata.previous-versions-max` | Integer | 100 | The max number of previous version metadata files to keep before deleting after commit. | +| Key | Options | Default | Description | +|------------------------------------------|-----------------------------------|---------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| +| `write.parquet.compression-codec` | `{uncompressed,zstd,gzip,snappy}` | zstd | Sets the Parquet compression coddec. | +| `write.parquet.compression-level` | Integer | null | Parquet compression level for the codec. If not set, it is up to PyIceberg | +| `write.parquet.row-group-limit` | Number of rows | 1048576 | The upper bound of the number of entries within a single row group | +| `write.parquet.page-size-bytes` | Size in bytes | 1MB | Set a target threshold for the approximate encoded size of data pages within a column chunk | +| `write.parquet.page-row-limit` | Number of rows | 20000 | Set a target threshold for the maximum number of rows within a column chunk | +| `write.parquet.dict-size-bytes` | Size in bytes | 2MB | Set the dictionary page size limit per row group | +| `write.metadata.previous-versions-max` | Integer | 100 | The max number of previous version metadata files to keep before deleting after commit. | +| `write.object-storage.enabled` | Boolean | True | Enables the [`ObjectStoreLocationProvider`](configuration.md#object-store-location-provider) that adds a hash component to file paths. Note: the default value of `True` differs from Iceberg's Java implementation | +| `write.object-storage.partitioned-paths` | Boolean | True | Controls whether [partition values are included in file paths](configuration.md#partition-exclusion) when object storage is enabled | +| `write.py-location-provider.impl` | String of form `module.ClassName` | null | Optional, [custom `LocationProvider`](configuration.md#loading-a-custom-location-provider) implementation | ### Table behavior options @@ -195,6 +198,93 @@ PyIceberg uses [S3FileSystem](https://arrow.apache.org/docs/python/generated/pya +## Location Providers + +Apache Iceberg uses the concept of a `LocationProvider` to manage file paths for a table's data. In PyIceberg, the +`LocationProvider` module is designed to be pluggable, allowing customization for specific use cases. The +`LocationProvider` for a table can be specified through table properties. + +PyIceberg defaults to the [`ObjectStoreLocationProvider`](configuration.md#object-store-location-provider), which generates +file paths that are optimized for object storage. + +### Simple Location Provider + +The `SimpleLocationProvider` places a table's file names underneath a `data` directory in the table's base storage +location (this is `table.metadata.location` - see the [Iceberg table specification](https://iceberg.apache.org/spec/#table-metadata)). +For example, a non-partitioned table might have a data file with location: + +```txt +s3://bucket/ns/table/data/0000-0-5affc076-96a4-48f2-9cd2-d5efbc9f0c94-00001.parquet +``` + +When the table is partitioned, files under a given partition are grouped into a subdirectory, with that partition key +and value as the directory name - this is known as the *Hive-style* partition path format. For example, a table +partitioned over a string column `category` might have a data file with location: + +```txt +s3://bucket/ns/table/data/category=orders/0000-0-5affc076-96a4-48f2-9cd2-d5efbc9f0c94-00001.parquet +``` + +The `SimpleLocationProvider` is enabled for a table by explicitly setting its `write.object-storage.enabled` table +property to `False`. + +### Object Store Location Provider + +PyIceberg offers the `ObjectStoreLocationProvider`, and an optional [partition-exclusion](configuration.md#partition-exclusion) +optimization, designed for tables stored in object storage. For additional context and motivation concerning these configurations, +see their [documentation for Iceberg's Java implementation](https://iceberg.apache.org/docs/latest/aws/#object-store-file-layout). + +When several files are stored under the same prefix, cloud object stores such as S3 often [throttle requests on prefixes](https://repost.aws/knowledge-center/http-5xx-errors-s3), +resulting in slowdowns. The `ObjectStoreLocationProvider` counteracts this by injecting deterministic hashes, in the form of binary directories, +into file paths, to distribute files across a larger number of object store prefixes. + +Paths still contain partitions just before the file name, in Hive-style, and a `data` directory beneath the table's location, +in a similar manner to the [`SimpleLocationProvider`](configuration.md#simple-location-provider). For example, a table +partitioned over a string column `category` might have a data file with location: (note the additional binary directories) + +```txt +s3://bucket/ns/table/data/0101/0110/1001/10110010/category=orders/0000-0-5affc076-96a4-48f2-9cd2-d5efbc9f0c94-00001.parquet +``` + +The `write.object-storage.enabled` table property determines whether the `ObjectStoreLocationProvider` is enabled for a +table. It is used by default. + +#### Partition Exclusion + +When the `ObjectStoreLocationProvider` is used, the table property `write.object-storage.partitioned-paths`, which +defaults to `True`, can be set to `False` as an additional optimization for object stores. This omits partition keys and +values from data file paths *entirely* to further reduce key size. With it disabled, the same data file above would +instead be written to: (note the absence of `category=orders`) + +```txt +s3://bucket/ns/table/data/1101/0100/1011/00111010-00000-0-5affc076-96a4-48f2-9cd2-d5efbc9f0c94-00001.parquet +``` + +### Loading a Custom Location Provider + +Similar to FileIO, a custom `LocationProvider` may be provided for a table by concretely subclassing the abstract base +class [`LocationProvider`](../reference/pyiceberg/table/locations/#pyiceberg.table.locations.LocationProvider). + +The table property `write.py-location-provider.impl` should be set to the fully-qualified name of the custom +`LocationProvider` (i.e. `mymodule.MyLocationProvider`). Recall that a `LocationProvider` is configured per-table, +permitting different location provision for different tables. Note also that Iceberg's Java implementation uses a +different table property, `write.location-provider.impl`, for custom Java implementations. + +An example, custom `LocationProvider` implementation is shown below. + +```py +import uuid + +class UUIDLocationProvider(LocationProvider): + def __init__(self, table_location: str, table_properties: Properties): + super().__init__(table_location, table_properties) + + def new_data_location(self, data_file_name: str, partition_key: Optional[PartitionKey] = None) -> str: + # Can use any custom method to generate a file path given the partitioning information and file name + prefix = f"{self.table_location}/{uuid.uuid4()}" + return f"{prefix}/{partition_key.to_path()}/{data_file_name}" if partition_key else f"{prefix}/{data_file_name}" +``` + ## Catalogs PyIceberg currently has native catalog type support for REST, SQL, Hive, Glue and DynamoDB. diff --git a/pyiceberg/table/locations.py b/pyiceberg/table/locations.py index 046ee32527..53b41d1e61 100644 --- a/pyiceberg/table/locations.py +++ b/pyiceberg/table/locations.py @@ -30,7 +30,12 @@ class LocationProvider(ABC): - """A base class for location providers, that provide data file locations for write tasks.""" + """A base class for location providers, that provide data file locations for a table's write tasks. + + Args: + table_location (str): The table's base storage location. + table_properties (Properties): The table's properties. + """ table_location: str table_properties: Properties