Concurrent writes failures

[bikeshedder]

Apache Iceberg version

0.7.1

Please describe the bug 🐞

Summary

I'm currently trying to migrate a couple of dataframes with a custom hive-like storage scheme to Iceberg. After a lot of fiddling I managed to load the dataframes from an Azure storage, create the table in the Iceberg catalog (currently using sqlite + local fs) and append fragments from the Parquet dataset. As soon as adding a thread pool I always run into concurrency issues.

Errors

I get either of the following two error messages:

CommitFailedException: Requirement failed: branch main has changed: expected id 7548527194257629329, found 
8136001929437813453

or

CommitFailedException: Requirement failed: branch main was created concurrently

Sources

I use Dataset.get_fragments and insert the data into an iceberg table with identical partitioning.

I can work around this error by using a GIL (global iceberg lock, pun intended.) which is just a threading.Lock() that ensures every load_table() + table.append happens atomically. But that kills almost all performance gains there could be made. Also I plan on using this in some Celery runners . So using a threading.Lock() is no option in the future anyways.

azure_import.py

#!/bin/env -S poetry run python

from concurrent.futures import ThreadPoolExecutor, as_completed

import pyarrow as pa
import pyarrow.dataset as pd
from adlfs import AzureBlobFileSystem
from azure.identity import DefaultAzureCredential
from azure.storage.blob import BlobServiceClient
from pyarrow.dataset import HivePartitioning
from pyiceberg.catalog import Catalog
from pyiceberg.catalog.sql import SqlCatalog
from pyiceberg.io.pyarrow import pyarrow_to_schema
from pyiceberg.partitioning import PartitionField, PartitionSpec
from pyiceberg.table.name_mapping import MappedField, NameMapping
from pyiceberg.transforms import IdentityTransform

import settings


class AzureStorage:
    def __init__(self):
        credential = DefaultAzureCredential()
        blob_service_client = BlobServiceClient(
            settings.AZURE_BLOB_URL, credential
        )
        self.container_client = blob_service_client.get_container_client(
            settings.AZURE_BLOB_CONTAINER
        )
        # The AzureBlobFileSystem doesn't cleanly shutdown and currently
        # always raises an expection at the end of this program. See:
        # https://github.com/fsspec/adlfs/issues/431
        self.abfs = AzureBlobFileSystem(
            account_name=settings.AZURE_BLOB_ACCOUNT_NAME,
            credential=credential,
        )

    def list_tables(self):
        return self.container_client.walk_blobs(
            settings.AZURE_LIVE_PATH, delimiter="/"
        )

    def load_dataset(self, table_name) -> pd.Dataset:
        name = "/".join((settings.AZURE_LIVE_PATH.rstrip("/"), table_name))
        dataset = pd.dataset(
            "/".join([settings.AZURE_LIVE_CONTAINER, name]),
            format="parquet",
            filesystem=self.abfs,
            partitioning=HivePartitioning(
                pa.schema(
                    [
                        ("dataset", pa.string()),
                        ("flavor", pa.string()),
                    ]
                )
            ),
        )
        return dataset


def create_iceberg_catalog():
    catalog = SqlCatalog(
        "default",
        **{
            "uri": settings.ICEBERG_DATABASE_URI,
            "warehouse": settings.ICEBERG_WAREHOUSE,
        },
    )
    return catalog


def download_table(catalog: Catalog, table_name: str, ds: pd.Dataset):
    name_mapping = NameMapping(
        root=[
            MappedField(field_id=field_id, names=[field.name])
            for field_id, field in enumerate(ds.schema, 1)
        ]
    )
    schema = pyarrow_to_schema(ds.schema, name_mapping=name_mapping)
    assert isinstance(ds.partitioning, HivePartitioning), ds.partitioning
    partitioning_spec = PartitionSpec(
        *(
            PartitionField(
                source_id=name_mapping.find(field.name).field_id,
                field_id=-1,
                transform=IdentityTransform(),
                name=field.name,
            )
            for field in ds.partitioning.schema
        )
    )
    table = catalog.create_table(
        f"{settings.ICEBERG_NAMESPACE}.{table_name}",
        schema=schema,
        partition_spec=partitioning_spec,
    )
    fragments = list(ds.get_fragments())
    with ThreadPoolExecutor(8) as executor:
        futures = [
            executor.submit(
                download_fragment,
                table.identifier,
                fragment,
            )
            for fragment in fragments
        ]
        for future in as_completed(futures):
            try:
                future.result()
            except Exception as e:
                executor.shutdown(wait=False, cancel_futures=True)
                raise e from None


def download_fragment(
    table_identifier: str,
    fragment,
):
    catalog = create_iceberg_catalog()
    partition_keys = pd.get_partition_keys(fragment.partition_expression)
    fragment_table = fragment.to_table()
    for k, v in partition_keys.items():
        fragment_table = fragment_table.append_column(
            pa.field(k, pa.string(), nullable=False),
            pa.repeat(pa.scalar(v), fragment_table.num_rows),
        )
    table = catalog.load_table(table_identifier)
    table.append(fragment_table)


def import_data(storage: AzureStorage, catalog, table_name):
    dataset = storage.load_dataset(table_name)
    download_table(catalog, table_name, dataset)


def main():
    catalog = create_iceberg_catalog()
    catalog.create_namespace_if_not_exists(settings.ICEBERG_NAMESPACE)
    storage = AzureStorage()
    for table_name in storage.list_tables():
        import_data(storage, catalog, table_name)


if __name__ == "__main__":
    main()

pyproject.toml

[tool.poetry]
name = "iceberg-azure-importer"
version = "0.1.0"
description = ""
authors = ["Michael P. Jung <michael.jung@terreon.de>"]
package-mode = false

[tool.poetry.dependencies]
python = "^3.12"
pyiceberg = { extras = ["sql-postgres"], version = "^0.7.1" }
azure-identity = "^1.17.1"
adlfs = "^2024"
psutil = "^6.0.0"
pyarrow = "^17.0.0"
fsspec = "^2024"

[bikeshedder]

As pointed out in the Slack channel by @sungwy this is caused by the following two issues:

• Support intelligent commit retries #269
• Support IsolationLevels and Concurrency Safety Validation Checks #819

[TiansuYu]

^ Second link should be this one #819

[kevinjqliu]

currently using sqlite + local fs

FYI, according to the docs, "SQLite is not built for concurrency, you should use this catalog for exploratory or development purposes."
https://py.iceberg.apache.org/configuration/#sql-catalog

[bikeshedder]

FYI, according to the docs, "SQLite is not built for concurrency, you should use this catalog for exploratory or development purposes." https://py.iceberg.apache.org/configuration/#sql-catalog

I know. This issue exists with both PostgreSQL and SQLite. SQLite just makes the reproduction a bit simpler. You're right pointing it out though. Other users might want to use SQLite in production otherwise.

[reinthal]

Here's some code that worked for me for me

def append_to_table_with_retry(pa_df: pa.Table, table_name: str, catalog: Catalog) -> None:
    """Appends a pyarrow dataframe to the table in the catalog using tenacity exponential backoff."""
    @retry(
        wait=wait_exponential(multiplier=1, min=4, max=32),
        stop=stop_after_attempt(20),
        reraise=True
    )
    def append_with_retry():
        table = catalog.load_table(table_name) # <---- If a process appends between this line ...
        table.append(pa_df) # <----- and this line, then Tenacity will retry.

    append_with_retry()

[bikeshedder]

This doesn't work - at least not efficiently - if you're writing rather large files with a high concurrency.

e.g. Many threads uploading a 1 GB dataframe each can end up uploading every dataframe many times with this approach as it retries the entire operation. This is just a huge waste of bandwidth and performance and performs worse than implementing a GIL (Global Iceberg Lock).

I ended up migrating our data to ClickHouse. It's an entirely different beast but provides way better performance for our use case anyways. I'm happy to revisit pyiceberg once commit retries are implemented.