[Java] Provide light-weight arrow APIs

[asfimport]

We are trying to incorporate Apache Arrow to Apache Flink runtime. We find Arrow an amazing library, which greatly simplifies the support of columnar data format.

However, for many scenarios, we find the performance unacceptable. Our investigation shows the reason is that, there are too many redundant checks and computations in Arrow API.

For example, the following figures shows that in a single call to Float8Vector.get(int) method (this is one of the most frequently used APIs in Flink computation),  there are 20+ method invocations.

 

There are many other APIs with similar problems. We believe that these checks will make sure of the integrity of the program. However, it also impacts performance severely. For our evaluation, the performance may degrade by two or three orders of magnitude slower, compared to access data on heap memory. 

We think at least for some scenarios, we can give the responsibility of integrity check to application owners. If they can be sure all the checks have been passed, we can provide some light-weight APIs and the inherent high performance, to them.

In the light-weight APIs, we only provide minimum checks, or avoid checks at all. The application owner can still develop and debug their code using the original heavy-weight APIs. Once all bugs have been fixed, they can switch to light-weight APIs in their products and enjoy the consequent high performance.

 

Reporter: Liya Fan / @liyafan82
Assignee: Liya Fan / @liyafan82

Related issues:

• [Java] Add performance benchmarks from SQL workloads (is blocked by)
• [Java] Add microbenchmark code to Java (relates to)
• [Java] Improving Arrow Vector Reading performance (relates to)
• [Java] Add accessor methods for data buffers that skip null checking (relates to)
• [Java] Provide a flag to enable/disable null-checking in vectors' get methods (is related to)

Original Issue Attachments:

• image-2019-04-23-15-19-34-187.png
• safe_nocheck.jpg
• unsafe.jpg

PRs and other links:

• GitHub Pull Request apache/arrow#4212

_{Note: This issue was originally created as ARROW-5200. Please see the migration documentation for further details.}

[asfimport]

Wes McKinney / @wesm:
This is related to at least a few other JIRAs. It would be great to start to set up some microbenchmarks so that it is possible to quantify the improvements in various use cases, otherwise we're shooting in the dark. I think having additional "unsafe" methods that skip checks that prevent Java off-heap memory-related segfaults is a good idea

[asfimport]

Liya Fan / @liyafan82:
Sounds reasonable. Thanks a lot for your comments. 

We have opened a new Jira (ARROW-5209) to setup some performance benchmarks from our SQL engine, which is going to be made open source. The benchmarks are extracted by running an open SQL benchmark TPC-H. 

[asfimport]

Liya Fan / @liyafan82:
This is the (source/byte code/assembly) code generated by the original Arrow API for Float8Vector.

And this is the code generated by the unsafe API for Float8Vector.

It can be observed that the amount of (source/byte code/assembly) code generated by unsafe API is smaller.

[asfimport]

Micah Kornfield / @emkornfield:
[~fan_li_ya] I think this can be closed as won't fix and so can the corresponding pull request?

[asfimport]

Liya Fan / @liyafan82:
[~[email protected]] Closed. Thanks for your kind reminder.