Pietro Cerutti created AVRO-3710:
------------------------------------
Summary: C++ - don't take ownership of I/OStream in DataFile
Key: AVRO-3710
URL: https://issues.apache.org/jira/browse/AVRO-3710
Project: Apache Avro
Issue Type: Improvement
Components: c++
Reporter: Pietro Cerutti
In AVRO-2014, I raised concerns regarding how DataFile(Reader|Writer)[Base]
take their streams by unique_ptr.
Here, I would like to propose a fix.
The problem is that, because streams are taken by unique_ptr, caller code
doesn't have access to the streams after the construction of the DataFile*
object. This makes it impossible to use custom streams, or even the built-in
memory streams:
{code:cpp}
auto schema{ giveMeMySchema(); }
auto os{ avro::memoryOutputStream() };
DataFileWriter w{ os, schema };
auto is{ memoryInputStream(*os) }; // ouch: os is gone
doSomethingWithTheAvroStream(is);
{code}
I am proposing an almost backwards-compatible change, which is to change the
DataFile classes to take and hold the streams by shared_ptr.
The semantics for client code don't change: you can still move a
unique_ptr<Stream> into the DataFile constructor, and in that case the DataFile
will be the only owner.
But this enable a client from doing something like this:
{code:cpp}
auto schema{ giveMeMySchema(); }
auto os{ avro::memoryOutputStream() };
std::shared_ptr<OutputStream> shared_os{ os.get(), boost::null_deleter{} };
DataFileWriter w{ shared_os, schema };
auto is{ memoryInputStream(*os) }; // good: os is alive
doSomethingWithTheAvroStream(is);
{code}
I will submit a PR if this is accepted
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
