avro-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From John Lilley <john.lil...@redpoint.net>
Subject RE: State of the C++ vs Java implementations
Date Mon, 18 Aug 2014 12:53:19 GMT
Thanks Steve!

As for this approach: "As for the compression - as Doug has already answered, C++ only supports
null (no) codec and deflate. You can always use the Avro java tools 'recodec' command to convert
from an unsupported codec to deflate if you need to. "

It's not an option for us.  We want to read Avro files natively, as quickly as possible. 
Also it doesn't look like these tools read/write HDFS paths do they?

john


From: Steve Roehrs [mailto:Steve.Roehrs@rlmgroup.com.au]
Sent: Sunday, August 17, 2014 6:41 PM
To: John Lilley; user@avro.apache.org
Subject: RE: State of the C++ vs Java implementations

Hi John.

Sorry for the late reply I was off work for a few days ill.

The idea of using reading the schema from the file and then processing it without knowing
the structure beforehand is the main use case for the GenericDatum.  The inefficiencies relate
to the way that GenericDatum handles arrays.  In our use case, most of the data consists of
large floating point arrays, or multi-dimensional (nested) arrays.  GenericDatum stores these
by using a STL vector.

So if you were using C++ structures, you may have a float[1000] - but if you use GenericDatum
you get a  vector<GenericDatum> where each GenericDatum contains a single float.  Of
course this is the most flexible way of implementing it, and it works - but it uses considerably
more memory.  Profiling a read of one of our data structures showed that more than 50% of
the time was spent in malloc()/ free() !

If your data doesn't have lots of large numeric arrays then by all means, GenericDatum should
work reasonably well for you.

As for the compression - as Doug has already answered, C++ only supports null (no) codec and
deflate. You can always use the Avro java tools 'recodec' command to convert from an unsupported
codec to deflate if you need to.

The groundwork for codec support in C++ is already there - it should be quite easy to add
additional codecs now. Most of the work would be in getting the makefile/library stuff right.

Regards,

Steve Roehrs
Senior Software Engineer | Lockheed Martin

| p: +61 8 7389 4525    | m: +61 4 3891 5622     | f: +61 8 7389 4551
| w: www.rlmgroup.com.au<http://www.rlmgroup.com.au> | e: Steve.Roehrs@rlmgroup.com.au<mailto:Steve.Roehrs@rlmgroup.com.au>
| Company address: 82-86 Woomera Ave, Edinburgh, SA 5111
This email and any attachment to it remains the property of Lockheed Martin and is intended
only to be read or used by the named addressee.  It may contain information that is confidential,
commercially valuable or subject to legal privilege.  If you receive this email in error,
please immediately delete it and notify the sender.  Opinions, conclusions and other information
in this message that do not relate to the official business of Lockheed Martin or any companies
within Lockheed Martin shall be understood as neither given nor endorsed by them.
________________________________
From: John Lilley [mailto:john.lilley@redpoint.net]
Sent: Friday, August 15, 2014 1:16 AM
To: user@avro.apache.org<mailto:user@avro.apache.org>; Steve Roehrs
Subject: RE: State of the C++ vs Java implementations

Steve,

Thanks so much for the reply.  I hope that I can inconvenience you for a little more guidance.
 We want to read and write Avro data files whose schema is not known until run-time, when
we read the file metadata and transform that into our own internal record structure.  So we
are not mapping to a C++ struct/class with defined compile-time members.  We just want to
loop over the records and columns in the data file, transforming them serially.  Can this
be done without incurring the performance penalty of GenericDatum that you speak of?

Different question: do you know if the full complement of compression codecs is available
in C++?  We don't need "everything possible", but we want to be able to read 99.9% of files
that we are likely to encounter in practice.

Thanks
John


From: Steve Roehrs [mailto:Steve.Roehrs@rlmgroup.com.au]
Sent: Sunday, August 10, 2014 11:25 PM
To: user@avro.apache.org<mailto:user@avro.apache.org>
Subject: RE: State of the C++ vs Java implementations

Hi John

You can definitely read and write Avro data files using C++.  The DataFileWriter and DataFileReader
classes are what you need.

The README is severely out of date.

I can't comment on the relative performance of the Java/C++ API's - we used the C++ API for
our application, but for performance reasons we don't use the GenericDatum class, as it does
have poor performance for our particular mix of data.  I don't know if the Java API fares
any better in this regard.

Regards,

Steve Roehrs
Senior Software Engineer | Lockheed Martin

| p: +61 8 7389 4525    | m: +61 4 3891 5622     | f: +61 8 7389 4551
| w: www.rlmgroup.com.au<http://www.rlmgroup.com.au> | e: Steve.Roehrs@rlmgroup.com.au<mailto:Steve.Roehrs@rlmgroup.com.au>
| Company address: 82-86 Woomera Ave, Edinburgh, SA 5111
This email and any attachment to it remains the property of Lockheed Martin and is intended
only to be read or used by the named addressee.  It may contain information that is confidential,
commercially valuable or subject to legal privilege.  If you receive this email in error,
please immediately delete it and notify the sender.  Opinions, conclusions and other information
in this message that do not relate to the official business of Lockheed Martin or any companies
within Lockheed Martin shall be understood as neither given nor endorsed by them.
________________________________
From: John Lilley [mailto:john.lilley@redpoint.net]
Sent: Wednesday, August 06, 2014 6:28 AM
To: user@avro.apache.org<mailto:user@avro.apache.org>
Subject: State of the C++ vs Java implementations

Greetings,

I am desiring to read and write Avro files (such as those manipulated by MapReduce applications)
from a C++ program.  While there are higher-level wrappers (such as Hive), I am interested
in reading/writing the files directly.  There are both C++ and Java library implementations;
however, in the C++ API README I see "And the file and rpc containers are not yet implemented."
 Does this mean that I can't read and write Avro files using the C++ library?

We have very good C++/JNI wrapper-generator, so using the Java is not terribly difficult.
 Given that, which interface would you recommend?  Does the C++ interface (assuming it works)
have significant performance advantages?

Thanks
john

Mime
View raw message