Tiago Margalho created AVRO-3223:
------------------------------------
Summary: Allow dynamic codec support
Key: AVRO-3223
URL: https://issues.apache.org/jira/browse/AVRO-3223
Project: Apache Avro
Issue Type: New Feature
Components: csharp
Reporter: Tiago Margalho
I'd like to propose a change to the C# library to allow it to dynamically
support codecs that aren't supported out of the box. That is, instead of adding
more codecs to the library, the library should allow codecs to be resolved by
the user. At the moment, only deflate is supported by the library.
In fact, the library already supports "custom" codecs to be used when writing
to an Avro file. A user can implement a Codec and pass it into the call to
OpenWriter:
{code:c#|title=DataFileWriter.cs|borderStyle=solid}
public static IFileWriter<T> OpenWriter(DatumWriter<T> writer, string path,
Codec codec);
{code}
However, there's no way a user can call DataFileReader.OpenRead or
DataFileReader.OpenAppendWriter and specify how the codec should be resolved.
So what I'm proposing is that we allow these other methods to resolve custom
codecs thus enabling the library to support codecs that are currently supported
in other languages but not currently on the C# implementation.
Here are some reasons why I believe this would be a good approach to extend
support to other codecs:
- The .NET base class library may not support all codecs.
- The library won't have to depend on other OSS projects with potentially
conflicting license agreements.
- The user is allowed to pick which implementation of zstd, snappy, etc is
best for their use case (eg. may prefer a version compatible to an older
version of .NET, or one that is more recent and optimised for more recent
versions of .NET/C#).
h2. How this would work
At present, the library uses a codec resolution method that is only internally
used:
{code:c#|title=DataFileReader.cs|borderStyle=solid}
private Codec ResolveCodec()
{
return
Codec.CreateCodecFromString(GetMetaString(DataFileConstants.MetaDataCodec));
}
{code}
This can be somewhat easily extended to support other "resolvers" but first we
would need to allow "resolvers" to be registered to be discovered by the
DataFileReader:
{code:c#|title=DataFileReader.cs|borderStyle=solid}
/// <summary>
/// Represents a function capable of resolving a codec identifying
string
/// into a matching codec implementation a reader can use to decompress
data.
/// </summary>
public delegate Codec CodecResolver(string codecMetaString);
private static List<CodecResolver> _codecResolvers = new
List<CodecResolver>();
/// <summary>
/// Registers a function that will attempt to resolve a codec
identifying string
/// into a matching codec implementation when reading compressed Avro
data.
/// </summary>
public static void RegisterCodecResolver(CodecResolver resolver)
{
_codecResolvers.Add(resolver);
}
{code}
Thus, the implementation of `ResolveCodec` could be:
{code:c#|title=DataFileReader.cs|borderStyle=solid}
private Codec ResolveCodec()
{
var codecString = GetMetaString(DataFileConstants.MetaDataCodec);
foreach (var resolver in _codecResolvers)
{
var candidateCodec = resolver(codecString);
if (candidateCodec != null)
{
return candidateCodec;
}
}
return Codec.CreateCodecFromString(codecString);
}
{code}
h2. One additional change required
The only change left to make is related to how decompression buffers are passed
around. When I tried adding support for zstd using this approach, I found that
I couldn't decompress data because the DataFileReader reuses a `DataBlock`
buffer and the actual compressed data length doesn't necessarily match the
length of the buffer. Although some compression algos are likely to have
metadata in frames/headers describing the length of the encoded payload, not
all implementations will allow the buffer size to be different so I think it
would make sense to make the buffer size explicit.
As a consequence I had to make a change that allows us to pass the length of
the data in the DataBlock into the Codec.Decompress method call:
{code:c#|title=DataFileReader.cs|borderStyle=solid}
if (HasNextBlock())
{
_currentBlock = NextRawBlock(_currentBlock);
_currentBlock.Data = _codec.Decompress(_currentBlock.Data,
(int)this._blockSize);
...
{code}
I will be raising a PR shortly with the changes I've described above. Hopefully
people will find this proposal reasonable but please let me know what you think!
Thanks!
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
