schema-sanitizer

Version 0.3.5: this project is still being tuned and tested, especially for generating Parquet files used by BigQuery external tables.

schema-sanitizer converts messy CSV, JSON, JSON Lines, NDJSON, XML, and Parquet data into stable analytical tables or sanitized files. The native C++23 core handles schema inference, scalar/container reconciliation, field versioning, bounded streaming, and Arrow C Data materialization.

Index

• Install
• Public API
• Input Formats
• Input Mode
• Shared Parameters
• Paths And Input Selection
• Schema And Field Handling
• String Scalar Parsing
• Source-Specific Parsing
• Errors And Resources
• Configuration Examples
• Result
• ETL Generated Columns
• Registry Probes
• Schema Reconciliation
• Field Names
• Timestamp Precision
• Depth Limits
• Memory Safety And Tuning
• Filesystems
• Example 7
• Development
• License

Install

pip install 'schema-sanitizer[pyarrow]'

Optional analytical targets:

pip install 'schema-sanitizer[pandas]'
pip install 'schema-sanitizer[polars]'
pip install 'schema-sanitizer[duckdb]'
pip install 'schema-sanitizer[cloud]'
pip install 'schema-sanitizer[all]'

import schema_sanitizer as ss

Public API

All public operations are named to_*.

In-memory analytical functions:

File-to-file functions:

events = ss.to_pyarrow(
    "raw/events.jsonl",
    input_format="jsonl",
)

customers = ss.to_pandas(
    "raw/customers.csv",
    input_format="csv",
)

ss.to_parquet(
    "raw/events.jsonl",
    "silver/events.parquet",
    input_format="jsonl",
)

All seven functions expose the same input and cleaning options. File-to-file functions additionally take output_path. to_parquet also accepts Parquet output compression options.

Input Formats

input_format must always be selected explicitly. The signature default is None, but calling any to_* function with None raises an error. Neither None nor "auto" infers a format from the extension or file contents.

The selected format also validates the source extension. For .json files, choose "json" for one document treated as one row or "json_array" for a top-level array of row objects.

jsonl and ndjson use the same newline-delimited JSON parser. Their only difference is the required extension.

Valid JSONL or NDJSON:

{"a": 1}
{"a": 2}

Valid json_array:

[
  {"id": 1, "name": "Ana"},
  {"id": 2, "name": "Luis"},
  {"id": 3, "name": "Marta"}
]

Every top-level json_array element must be an object. The array is split incrementally into rows instead of being materialized as one nested value.

Passing a mismatched extension fails before ingestion:

# Raises: jsonl requires .jsonl, not .ndjson
ss.to_pyarrow("events.ndjson", input_format="jsonl")

Input Mode

input_mode accepts:

Directory traversal is non-recursive. Files with other extensions and nested directories are ignored.

table = ss.to_pyarrow(
    "raw/2026-01/",
    input_format="jsonl",
    input_mode="directory",
).clean_data

Directory behavior:

• jsonl reads only direct .jsonl children.
• ndjson reads only direct .ndjson children.
• json reads direct .json documents as rows.
• json_array flattens each direct .json array into rows.
• csv removes repeated matching headers and rejects header mismatches.
• xml combines direct .xml documents and requires a compatible root/row tag.
• parquet streams direct .parquet and .pq children.

Directory mode requires an explicit input_format.

Shared Parameters

result = ss.to_pyarrow(
    input_path,
    input_format="jsonl",
    input_mode="single_file",
    schema_mode="additive",
    column_order="alphabetically",
    field_name_policy="lower_alpha",
    timestamp_precision="TIMESTAMP_MICROS",
    parse_integers=False,
    parse_floats=False,
    parse_float_decimal_separator=".",
    parse_float_thousands_separator=",",
    parse_iso_timestamps=False,
    parse_iso_dates=False,
    parse_iso_times=False,
    true_tokens=(),
    false_tokens=(),
    custom_timestamp_patterns=(),
    custom_date_patterns=(),
    custom_time_patterns=(),
    arrow_max_depth=32,
    parquet_max_depth=15,
    scalar_object_key="default_key",
    csv_has_header=True,
    csv_delimiter=",",
    input_text_encoding="utf-8",
    xml_row_tag=None,
    on_error="emit_null_row",
    batch_memory_limit_bytes=None,
    read_chunk_bytes=1024 * 1024,
    schema_registry=None,
)

Paths And Input Selection

Schema And Field Handling

With column_order="alphabetically", field ordering is recursive and canonical: root fields and every nested struct are sorted by sanitized output name. The same ordering is applied after additive schema evolution, so newly discovered columns can reshuffle previously registered columns inside their struct. This order is used for in-memory Arrow results, physical Parquet materialization, the stored schema registry canonical schema, and BigQuery external table DDL generated by the built-in BigQuery integration. Use column_order="schema_contract_first" only when preserving registered field order is more important than alphabetical layout.

String Scalar Parsing

These options apply to string values such as CSV cells, XML text, and quoted JSON values. Actual JSON numbers and booleans are already typed by JSON syntax and do not depend on these options.

When one field contains both integer and float values in the same inference batch, the field is inferred as float64; integer values are safely cast to float. During registry-backed incremental runs, existing float64 fields also accept later integer-only batches without creating an integer field version.

Source-Specific Parsing

Native path readers decode utf-8, utf-16, and latin-1/iso-8859-1 directly for CSV, JSON, JSON-array, and XML inputs, including directory and warm-up source plans. Other Python codecs keep the compatibility stream fallback.

Errors And Resources

Parquet Output

These options apply only to to_parquet.

Default GZIP output does not require passing either option:

ss.to_parquet(
    "raw/events.jsonl",
    "silver/events.parquet",
    input_format="jsonl",
)

Pass the options when you want to make the setting explicit or tune the zlib level:

ss.to_parquet(
    "raw/events.jsonl",
    "silver/events.parquet",
    input_format="jsonl",
    parquet_compression="gzip",
    parquet_gzip_level=6,
)

parquet_gzip_level accepts 0 through 9: lower values write faster with larger files; higher values spend more CPU to reduce file size. Leave it as None to use the writer/zlib default. For debugging or compatibility testing, disable compression explicitly:

ss.to_parquet(
    "raw/events.jsonl",
    "silver/events.uncompressed.parquet",
    input_format="jsonl",
    parquet_compression="uncompressed",
)

ISO timestamp, date, and time parsing is opt-in. With all three parse_iso_* flags left at False, ISO-looking source strings remain strings. The custom_*_patterns options are independent: configured custom patterns are still applied even when the corresponding built-in ISO parser is disabled.

Float separator options apply only when parse_floats=True and only to string values, including CSV cells and XML text. Real JSON numbers always use JSON's . decimal syntax. Grouping is strict: the default configuration accepts "1,234.56", while European input can use:

result = ss.to_pyarrow(
    "prices.csv",
    input_format="csv",
    parse_floats=True,
    parse_float_decimal_separator=",",
    parse_float_thousands_separator=".",
)

That configuration accepts "1.234,56" and "1234,56". Grouped sections after the first must contain exactly three digits. In comma-delimited CSV, values containing commas must be quoted.

Enabled string-to-scalar parsers first test the source string unchanged. If that strict attempt fails, they retry once after removing surrounding ASCII spaces, tabs, line breaks, form feeds, and vertical tabs. This applies to integer, float, Boolean-token, ISO temporal, and custom temporal parsing. The retry does not allocate or modify the source value:

" 123456"       -> 123456       when parse_integers=True
" yes "         -> true         when "yes" is a true token
" 2026-01-02 "  -> 2026-01-02   when parse_iso_dates=True

An unmatched string retains its original whitespace, and a whitespace-only string remains a string. Exact configured values are tested before trimming, so custom tokens or temporal patterns that intentionally include surrounding whitespace continue to work.

Configuration Examples

European numeric and semicolon-delimited CSV:

prices = ss.to_pyarrow(
    "prices.csv",
    input_format="csv",
    csv_delimiter=";",
    parse_floats=True,
    parse_float_decimal_separator=",",
    parse_float_thousands_separator=".",
).clean_data

Custom Boolean and temporal strings:

events = ss.to_pandas(
    "events.ndjson",
    input_format="ndjson",
    true_tokens=("yes", "active"),
    false_tokens=("no", "inactive"),
    parse_iso_timestamps=True,
    custom_date_patterns=(r"(\d{4})-(\d{2})-(\d{2})",),
).clean_data

Strict incremental conversion using an existing registry:

result = ss.to_parquet(
    "raw/events.jsonl",
    "silver/events.parquet",
    input_format="jsonl",
    schema_mode="strict",
    schema_registry=previous_result.schema_registry,
    on_error="stop",
)

Memory-first processing of a large directory:

result = ss.to_parquet(
    "raw/2026-01/",
    "silver/2026-01.parquet",
    input_format="jsonl",
    input_mode="directory",
    batch_memory_limit_bytes=64 * 1024 * 1024,
    read_chunk_bytes=256 * 1024,
)

to_csv, to_jsonl, and to_parquet return Result.clean_data is None. Analytical functions return their named in-memory object.

Result

Every public function returns schema_sanitizer.Result.

Analytical and file outputs use the same registry-backed native path, so they produce the same schema and metadata behavior.

ETL Generated Columns

Every analytical and file conversion adds these fixed top-level columns:

source_file and ingestion_timestamp contain values in every output row. schema_registry and schema_drifts contain values only in the first output row; remaining rows are null to avoid repeating large registry payloads.

Generic schema_registry value:

{
  "field_name_policy": "lower_snake",
  "registry_version": 1,
  "schema_generation": 2,
  "canonical_schema": {
    "fields": [
      {
        "name": "amount",
        "nullable": true,
        "type": {"kind": "string"}
      }
    ]
  },
  "variants": {
    "amount": {
      "versions": [
        {
          "output_name": "amount",
          "schema": "string",
          "is_most_compatible_current_version": true
        }
      ]
    }
  }
}

Generic schema_drifts value after amount acquires a float alternative:

[
  {
    "detected_at": "2026-06-25T09:05:08.947122Z",
    "source_path": "amount",
    "output_name": "amount_v2_float",
    "drift_type": "new_version_generated",
    "previous_schema": "string",
    "new_schema": "double"
  }
]

The names are part of the ETL output contract and cannot be configured. They are reserved at the top level: conversion fails before writing if the source schema already contains any of them. A nested source key such as payload.source_file is allowed because it does not conflict with the generated top-level columns; normal field-name sanitization still applies to that nested key. Reserved top-level source fields are not renamed or versioned, since silently doing so would make downstream registry discovery ambiguous.

Registry Probes

A registry probe is a schema-inference pass that reads source data and returns an updated schema_registry without writing the converted rows to the final output. Probes are used when the library needs the schema contract before the real materialization pass, for example:

• warm-up ranges in the pipeline;
• directory or multi-source inputs that must share one schema;
• registry-backed strict attempts before additive fallback;
• Parquet direct-ingest planning, where source Arrow schemas can be merged before a native output stream is opened.

Probes are not samples. When a probe is required, it scans every selected source file and every selected row needed for schema inference. Memory safety comes from the same bounded, replayable source planning used by normal conversion: remote objects are staged into local temporary files or chunks, directory children are processed incrementally, and native streams are closed after the probe result is produced.

The probe output is a registry state, not a data batch. The normal conversion then starts from that registry and processes its partition cleanly. Warm-up data is never appended to a normal partition Parquet file; it is used only to build or extend the initial schema contract.

When native support is available, probes carry a compiled native registry-state capsule alongside the JSON registry. Later probes and materialization can reuse that compiled state instead of reparsing the registry JSON, including Arrow-source and Parquet multi-source probes. The JSON schema_registry remains the durable public contract embedded in outputs and passed between processes; the native state is an in-process acceleration detail.

Probe and materialization paths intentionally share the same inference, versioning, field-name, depth, and scalar-parsing rules. This keeps warm-up, normal, analytical, and file-output behavior aligned: a field discovered by a probe is the same field the materializer will use, and drift discovered later is merged through the same registry reconciliation rules. For directory, remote-chunk, and Parquet multi-source inputs, normal file conversion also uses the same canonical native source-plan execution path where supported, so warm-up and normal runs do not maintain separate source-discovery or registry-stream heads.

Schema Reconciliation

The embedded schema_registry is the source of truth for incremental processing. Pass the latest registry to the next conversion:

result = ss.to_parquet(
    "raw/2026-01-09/events.jsonl",
    "silver/2026-01-09/events.parquet",
    input_format="jsonl",
    schema_registry=previous_registry,
)

next_registry = result.schema_registry

Before generating a field version, the native merge attempts compatible reconciliation:

• A singleton can be wrapped into an existing list.
• A scalar can be wrapped into an existing struct under default_key.
• Empty objects and lists provide no schema-inference evidence. If no other value or registry entry defines the field, the field is omitted. If the field is already established, the empty container materializes as null.
• New compatible struct children are added as nullable fields.

This rule applies recursively. Empty nested fields do not create child columns, affect sibling-name collision handling, trigger strict-schema extra-field errors, or generate schema drift. Empty elements inside an established list become null elements so list positions remain stable. Typed Parquet input keeps its declared columns on the direct Arrow path, but empty container values still become null and do not create additional type versions.

Irreconcilable drift creates a hybrid <original_name>_v<version>_<semantic_type> field at the lowest incompatible schema level. The original field remains unsuffixed and is version 1.

sentiment_analysis: struct<...>
sentiment_analysis_v2_struct_array: list<struct<
  magnitude: double,
  magnitude_v2_string: string
>>

The numeric component guarantees uniqueness and records discovery order within the registry. The semantic component describes the new logical type:

For example, list types produce integer_array, struct_array, or integer_array_array. If two incompatible alternatives have the same semantic type, their numeric versions still keep the columns distinct, such as payload_v2_struct_array and payload_v3_struct_array.

Existing exact historical variants are preferred during past-date reprocessing. Otherwise the newest compatible container is evolved recursively. Repeating an already known shape does not increment schema_generation.

Materialization routes each non-null source value to exactly one most-compatible member of its version family. It does not always choose the latest version:

• Arrays prefer list variants.
• Numeric values prefer numeric scalar variants.
• Ordinary strings prefer string variants.
• Parse-enabled numeric and temporal strings can target typed variants.
• A singleton can target a list variant and be wrapped as one element.
• Exact compatibility wins over fallback string conversion.
• If multiple versions receive the same compatibility score, the highest _vN_... version wins.
• A null source value leaves every member of the family null.

Given this family:

amount: string
amount_v2_integer: int64
amount_v3_float: double

values are routed as follows:

For a container family containing items: struct<...> and items_v2_struct_array: list<struct<...>>, both an array and a compatible singleton object go to items_v2_struct_array; the singleton is wrapped into a one-element list. Other family columns are null in that row.

Each drift event receives a native UTC detected_at timestamp. Row-level materialization time is available through ingestion_timestamp. Source partition identity remains available through source_file and any Hive partition columns, including during historical reprocessing.

Field Names

field_name_policy="lower_alpha" keeps lowercase a-z only. lower_snake keeps lowercase letters, digits, and underscores. preserve keeps source names.

Collisions use deterministic suffixes derived from the original dirty key, so source field order does not change the dirty-key to clean-key mapping.

Timestamp Precision

Accepted values:

• TIMESTAMP_MILLIS
• TIMESTAMP_MICROS (default)
• TIMESTAMP_NANOS

Microseconds are the default because BigQuery external tables support Parquet timestamp micros. BigQuery does not accept Parquet TIMESTAMP_NANOS.

Depth Limits

arrow_max_depth counts struct and list containers. parquet_max_depth counts Parquet/BigQuery RECORD levels; list wrappers do not add a RECORD level.

Over-depth nested values are flattened to string-compatible output rather than allowing unbounded schema expansion.

Memory Safety And Tuning

The pipeline uses replayable streaming sources, bounded inference batches, and streaming file writers. batch_memory_limit_bytes controls the approximate per-batch budget.

Memory-first settings for large files:

ss.to_parquet(
    "raw/large.jsonl",
    "silver/large.parquet",
    input_format="jsonl",
    batch_memory_limit_bytes=64 * 1024 * 1024,
    read_chunk_bytes=256 * 1024,
)

64 * 1024 * 1024 is 64 MiB.

Trade-offs:

• Lower batch_memory_limit_bytes reduces peak memory and may reduce speed.
• Lower read_chunk_bytes reduces transient input buffers and increases read calls.
• Parquet path decoding uses PyArrow's native dataset scanner and exports a bounded Arrow C stream to the sanitizer. This avoids Python per-batch ParquetFile.iter_batches iteration for local and staged path inputs while still scanning every row. Byte and file-like Parquet inputs keep the compatibility ParquetFile fallback.
• Parquet decoding enables threads only when the memory budget is large enough.
• Native Parquet output splits large Arrow batches into bounded row groups. The default native writer limit is 65,536 rows per row group; override it for a process with SCHEMA_SANITIZER_NATIVE_PARQUET_ROW_GROUP_ROWS=131072 when you want larger row groups, or a smaller value when limiting peak writer memory is more important. Row groups are also bounded by estimated uncompressed column bytes; override the 64 MiB default with SCHEMA_SANITIZER_NATIVE_PARQUET_ROW_GROUP_BYTES=33554432. When that byte limit is not explicitly set, the native writer lowers the effective default for very wide or repeated schemas so staging remains bounded for schema-sanitizer's nested outputs.
• Large native Parquet column chunks are split into bounded data pages. The default target is 1 MiB of staged uncompressed page data; override it with SCHEMA_SANITIZER_NATIVE_PARQUET_PAGE_BYTES=524288 when smaller pages are better for reader locality or memory.
• Native Parquet files include column and offset page indexes for flat columns. Readers that use Parquet page indexes can skip pages by statistics without scanning the full column chunk. Repeated/nested leaf columns intentionally do not get page indexes yet because their first-row indexes require extra repetition-level accounting.
• to_parquet uses parquet_compression="gzip" by default. Published wheels require zlib and smoke-test GZIP output in CI. Use parquet_compression="uncompressed" to disable compression or parquet_gzip_level=0..9 to tune the zlib level. The native writer still honors SCHEMA_SANITIZER_NATIVE_PARQUET_COMPRESSION and SCHEMA_SANITIZER_NATIVE_PARQUET_GZIP_LEVEL for low-level/internal calls, but public API parameters are preferred.
• Native Parquet output preserves supported Arrow integer widths, including int8, uint8, int16, uint16, uint32, and uint64, using Parquet integer logical annotations. It also writes timestamp millis, micros, and nanos with matching Parquet timestamp units. Schema-sanitizer all-null fields are written as Parquet Null; time32[s] fields are written as Parquet TIME(MILLIS) because Parquet has no seconds time unit.
• Native Parquet output chooses page encodings by compressed size. It uses dictionary encoding for profitable repeated scalar pages, DELTA_BINARY_PACKED for signed integer/date/time/timestamp pages when smaller, and DELTA_LENGTH_BYTE_ARRAY for high-cardinality string/binary pages when smaller. Non-dictionary float columns use Parquet BYTE_STREAM_SPLIT. Booleans and non-profitable pages stay plain-encoded. Known low-cardinality schema-sanitizer columns such as source_file, year, month, date, and hour prefer dictionary encoding on exact ties.
• Directory mode processes direct child files incrementally rather than loading the full directory at once.
• CSV directory normalization holds at most one configured-size source file in memory while validating and removing repeated headers.

Filesystems

Input and output paths may be local paths, file:// URIs, or supported cloud/object URIs such as:

file:///data/events.jsonl
s3://bucket/events/2026-01-09/events.jsonl
gs://bucket/events/2026-01-09/events.jsonl
https://storage.example.com/events/2026-01-09/events.jsonl
abfs://container@account.dfs.core.windows.net/events/2026-01-09/events.jsonl
https://account.blob.core.windows.net/container/events/2026-01-09/events.jsonl

Remote file I/O does not use pyarrow.fs. Public to_* calls stage remote inputs through provider-native async clients into replayable local temporary files, run the native sanitizer on those local paths, then upload file outputs to the requested remote destination. This avoids thousands of blocking remote opens and prevents the schema-inference/materialization passes from re-downloading the same source objects.

Single remote files are streamed into the local spool instead of being loaded as one in-memory byte payload. Remote directory children are fetched with a bounded concurrency window.

Supported remote backends:

Install cloud clients with:

pip install 'schema-sanitizer[cloud]'

Directory listing is non-recursive and deterministic. For remote directory inputs, matching child objects are listed asynchronously, downloaded with a bounded prefetch window, and written to a local spool before conversion:

• jsonl / ndjson: downloaded and concatenated with newline boundaries.
• json: each document is compacted to one JSONL row.
• json_array: each top-level array element becomes one JSONL row.
• csv: repeated matching headers are removed; mismatched headers fail.
• xml: documents are wrapped under one synthetic root after root-tag validation.
• parquet: child Parquet files are downloaded concurrently to a temporary local directory and then streamed through the existing Parquet reader.

Async remote I/O can be tuned with environment variables:

For very high counts of tiny source files, raise concurrency until the cloud service, local disk, or network saturates. Keep SCHEMA_SANITIZER_SPOOL_DIR on fast local storage with enough free space for one staged partition plus one staged output file. Registry-backed remote directory writes reuse the staged partition after schema probing, avoiding a second remote download pass.

Example 7

examples/example_07/07_gcs_jsonl_to_silver_parquet_range_prefix.py implements a single-writer GCS-to-Parquet pipeline with:

• CLI-selected input_format: csv, json, json_array, jsonl, ndjson, xml, or parquet
• CLI-selected input_mode: single_file or non-recursive directory
• daily year=YYYY/month=MM/date=YYYY-MM-DD partitions
• hourly year=YYYY/month=MM/date=YYYY-MM-DD/hour=HH partitions
• source extension validation derived from input_format
• integer, float, ISO timestamp, ISO date, and ISO time string parsing enabled
• source discovery and empty/missing partition skipping
• optional additive schema warm-up over a separate date/hour range before normal writes
• one sanitized Parquet output per logical partition
• embedded registry retrieval through Arrow ADBC
• optional native BigQuery registry sidecar table for latest-partition lookup
• incremental and random past-date reprocessing
• one final BigQuery external-table create/replace operation

In directory mode, all direct files matching input_format inside a source Hive partition are combined into that partition's single output Parquet. Subdirectories are not scanned.

Pipeline Flow

The example is intentionally single-writer. It builds the full logical partition plan first, performs async source discovery, removes missing or empty source partitions, optionally runs registry warm-up, writes selected Parquet partitions in order, and finally creates or replaces the BigQuery external table from the last output schema. The registry returned by each partition is carried into the next partition, so incremental schema state is monotonic within one process.

Source discovery uses the same local/cloud path logic for warm-up and normal runs. Remote directory discovery lists all matching objects for each selected partition and stages bounded chunks locally before native inference or conversion. When native support is available, remote normal conversion uses paired lazy chunk providers: one provider is consumed for full registry inference and a second provider streams the clean partition output with the final registry. The pipeline does not sample for schema inference: warm-up and normal additive inference scan all selected source files. Memory safety comes from bounded staged chunks, replayable temporary files, and native streaming rather than loading the full interval into memory.

Warm-Up Semantics

Use --start-date-warm-up and --end-date-warm-up to scan a registry warm-up range before writing normal outputs. Hour flags are valid only when --partition-granularity hourly is passed explicitly. In hourly mode, omitted normal hours default to the full day (0..23), and omitted warm-up hours also default to the full day when warm-up dates are present. Warm-up dates and hours are independent from the normal run range: they may be earlier, later, identical, or partially overlapping. Warm-up hour flags require both --start-date-warm-up and --end-date-warm-up.

The warm-up pass always merges schema additively, even when the normal run uses --schema-mode strict. All selected warm-up sources are treated as one logical inference input, not as independent partition outputs. If an existing registry was loaded from BigQuery first, warm-up starts from that registry and only adds new compatible fields or versions. If no existing registry is available, warm-up creates the initial registry.

When the warmed registry already covers a normal partition, that partition is written in a strict single pass. If a later normal partition introduces drift, the writer discards that temporary strict attempt and falls back to additive registry inference for that partition, then continues with the updated registry.

BigQuery Integration

BigQuery integration lives in schema_sanitizer.integrations.bigquery and uses Arrow ADBC. At startup, the example checks --target-table:

• if the table does not exist, additive mode or a warm-up range can bootstrap a fresh registry;
• if the table exists, it must be a BigQuery external table;
• configured Hive partition columns are validated against the existing table;
• the latest embedded schema_registry is read from the external table and used as the authoritative registry when it contains a canonical schema.

Without the sidecar feature, latest-registry lookup queries the external table for non-null schema_registry, ordered by ingestion timestamp, registry generation, and configured Hive partition columns. This can require scanning metadata across the external table, so large historical tables may be slow.

After successful writes, the example creates or replaces the external table DDL from the final Parquet schema. Hive partition fields are omitted from the Parquet-derived column list because BigQuery exposes them from the partitioned GCS path. The external-table spec is driven by:

• --external-table-source-uri, or derived <hive-prefix>/*;
• --external-table-hive-uri-prefix, or derived from the silver output prefix;
• --hive-partition-column name:TYPE, or daily/hourly defaults;
• --external-table-format, normally PARQUET;
• --external-table-require-partition-filter;
• --parquet-enable-list-inference;
• --parquet-compression, default gzip;
• --parquet-gzip-level, optional 0..9 gzip level.

BigQuery Registry Sidecar

Pass --bigquery-registry-sidecar-table project.dataset.table to maintain a native BigQuery sidecar table that avoids full external-table scans during registry bootstrap. The sidecar table has exactly two columns:

external_table_name STRING NOT NULL
last_ingested_partition STRING NOT NULL

external_table_name stores the referenced external table as project.dataset.table. last_ingested_partition stores the latest successful Hive partition key in the configured partition-column order:

year=2026/month=07/date=2026-07-05
year=2026/month=07/date=2026-07-05/hour=08

Custom partition columns use the same name=value/name=value encoding. The sidecar therefore works for daily, hourly, and user-configured Hive partition layouts without adding more columns.

When configured, registry bootstrap first reads the sidecar row for the target external table. If it finds a valid partition key, the embedded registry query is constrained to that one Hive partition. If the sidecar table is missing, is not a native table, has no row for the target external table, contains an invalid partition key, points to a partition without a valid registry, or the query fails, bootstrap falls back to the full external-table registry scan.

The sidecar is updated only after the selected Parquet outputs are written and the BigQuery external table has been created or replaced. The update is an idempotent BigQuery MERGE, so reruns and random past-date reprocessing update the pointer to the last partition completed by that run.

The partition control plane is available as reusable library code under schema_sanitizer.pipeline. It owns Hive date/hour URI planning, async source discovery, additive registry warm-up, and the registry-carrying partition write loop. It also exposes schema drift diff helpers used by the BigQuery example, reusable Parquet schema reads for local/remote outputs, and compact progress-log helpers. Cloud provider calls remain async Python I/O, while warm-up inference and conversion use the native registry-backed engine. Use schema_sanitizer.new_schema_registry() when a pipeline needs an empty registry document without depending on registry JSON internals. BigQuery external-table helpers live under schema_sanitizer.integrations.bigquery.

Daily single-file layout:

source-prefix/year=2026/month=06/date=2026-06-25/events_20260625.json
silver-prefix/year=2026/month=06/date=2026-06-25/events_20260625.parquet

Hourly directory layout:

source-prefix/year=2026/month=06/date=2026-06-25/hour=08/*.jsonl
silver-prefix/year=2026/month=06/date=2026-06-25/hour=08/events_20260625_08.parquet

For hourly layouts, pass --partition-granularity hourly. Use --start-hour and --end-hour to restrict the hourly partitions processed for every selected normal date; when they are omitted, hourly pipelines process hours 0..23.

Development

pip install -e .[dev]
pytest

Native build:

cmake -S . -B build/dev -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build/dev

License

Apache License 2.0. See LICENSE.