parquet | A library for reading and writing parquet files | Serialization library

It's not quite done yet, and it's not registered, but my rewrite of the Julia parquet package, Parquet2.jl does support both custom file metadata and individual column metadata (the keyword arguments metadata and column_metadata in Parquet2.writefile.

I haven't gotten to documentation for writing yet, but if you are feeling adventurous you can give it a shot. I do expect to finish up this package and register it within the next couple of weeks. I don't have unit tests in for writing yet, so of course, if you try it and have problems, please open an issue.

It's probably also worth mentioning that the main use case I recommend for parquet is if you must have parquet for compatibility reasons. Most of the time, Julia users are probably better off with Arrow.jl as the format has a number of advantages over parquet for most use cases, please see my FAQ answer on this. Of course, the reason I undertook writing the package is because parquet is arguably the only ubiquitous binary format in "big data world" so a robust writer is desperately needed.

This is possible but takes a little bit of work because in addition to being columnar Parquet also requires a schema.

The rough workflow is:

1. Open a parquet file for reading.

2. Then use iter_batches to read back chunks of rows incrementally (you can also pass specific columns you want to read from the file to save IO/CPU).

3. You can then transform each pa.RecordBatch from iter_batches further. Once you are done transforming the first batch you can get its schema and create a new ParquetWriter.

4. For each transformed batch call write_table. You have to first convert it to a pa.Table.

5. Close the files.

Parquet requires random access, so it can't be streamed easily from a URI (pyarrow should support it if you opened the file via HTTP FSSpec) but I think you might get blocked on writes.

I made it work by setting --conf to spark.sql.legacy.parquet.int96RebaseModeInRead=CORRECTED --conf spark.sql.legacy.parquet.int96RebaseModeInWrite=CORRECTED --conf spark.sql.legacy.parquet.datetimeRebaseModeInRead=CORRECTED --conf spark.sql.legacy.parquet.datetimeRebaseModeInWrite=CORRECTED.

This is a workaround though and Glue Dev team is working on a fix, although there is no ETA.

Also this is still very buggy. You can not call .show() on a DynamicFrame for example, you need to call it on a DataFrame. Also all my jobs failed where I call data_frame.rdd.isEmpty(), don't ask me why.

Update 24.11.2021: I reached out to the Glue Dev Team and they told me that this is the intended way of fixing it. There is a workaround that can be done inside of the script though:

This is possible by renaming the single parquet file in the dataset so that it always has the same file name, that way the export task will overwrite the previous file in the external system.

This can be done using raw file system access. The write_single_named_parquet_file function below validates its inputs, creates a file with a given name in the output dataset, then copies the file in the input dataset to it. The result is a schemaless output dataset that contains a single named parquet file.

Notes

• The build will fail if the input contains more than one parquet file, as pointed out in the question, calling .coalesce(1) (or .repartition(1)) is necessary in the upstream transform
• If you require transaction history in your external store, or your dataset is much larger than 512 MB this method is not appropriate, as only the latest version is kept, and you likely want multiple parquet files for use in your downstream system. The createTransactionFolders (put each new export in a different folder) and flagFile (create a flag file once all files have been written) options can be useful in this case.
• The transform does not require any spark executors, so it is possible to use @configure() to give it a driver only profile. Giving the driver additional memory should fix out of memory errors when working with larger datasets.
• shutil.copyfileobj is used because the 'files' that are opened are actually just file objects.

Full code snippet

example_transform.py

You can't change this behavior in the API, either when loading the parquet file into an arrow table or converting the arrow table to pandas.

But you can write your own function that would look at the schema of the arrow table and convert every list field to a python list

ABFS is a "real" file system, so the S3A zero rename committers are not needed. Indeed, they won't work. And the client is entirely open source - look into the hadoop-azure module.

the ADLS gen2 store does have scale problems, but unless you are trying to commit 10,000 files, or clean up massively deep directory trees -you won't hit these. If you do get error messages about Elliott to rename individual files and you are doing Jobs of that scale (a) talk to Microsoft about increasing your allocated capacity and (b) pick this up https://github.com/apache/hadoop/pull/2971

This isn't it. I would guess that actually you have multiple jobs writing to the same output path, and one is cleaning up while the other is setting up. In particular -they both seem to have a job ID of "0". Because of the same job ID is being used, what only as task set up and task cleanup getting mixed up, it is possible that when an job one commits it includes the output from job 2 from all task attempts which have successfully been committed.

I believe that this has been a known problem with spark standalone deployments, though I can't find a relevant JIRA. SPARK-24552 is close, but should have been fixed in your version. SPARK-33402 Jobs launched in same second have duplicate MapReduce JobIDs. That is about job IDs just coming from the system current time, not 0. But: you can try upgrading your spark version to see if it goes away.

My suggestions

1. make sure your jobs are not writing to the same table simultaneously. Things will get in a mess.
2. grab the most recent version spark you are happy with

I think you need give ParquetDataset a hint of the partition keys schema.

At least in arrow's implementation I would expect, but haven't verified the exact same input (including identical metadata) in the same order to yield deterministic outputs (we try not to leave uninitialized values for security reasons) with the same configuration (assuming the compression algorithm chosen also makes the deterministic guarantee). It is possible there is some hash-map iteration for metadata or elsewhere that might also break this assumption.

As @Pace pointed out I would not rely on this and recommend against relying on it). There is nothing in the spec that guarantees this and since the writer version is persisted when writing a file you are guaranteed a breakage if you ever decided to upgrade. Things will also break if additional metadata is added or removed ( I believe in the past there have been some big fixes for round tripping data sets that would have caused non-determinism).

So in summary this might or might not work today but even if it does I would expect this would be very brittle.

s3path and ids parameters that are passed to deleteIDs are not actually strings and sets respectively. They are instead columns.

In order to operate over these values you can instead create a UDF that accepts columns instead of intrinsic types, or you can collect your dataset if it is small enough so that you can use the values in the deleteIDs function directly. The former is likely your best bet if you seek to take advantage of Spark's parallelism.

You can read about UDFs here