How I Added Coprocessor ENUM/SET Support to TiKV in an LFX Mentorship Program

Xuanwo

March 30, 2021


In September 2020, I was honored to participate in the TiKV Linux Foundation’s (LFX) Mentorship Program. I worked on the Coprocessor support ENUM/SET project to help enable TiKV Coprocessor to support ENUM and SET calculations. This helped improve TiKV’s calculation performance.

In this post, I’ll share with you some background information about my project, how I implemented the project, my lessons learned during community cooperation, and my future plans for this project.

Background information

What is the LFX Mentorship Program?

The LFX Mentorship Programs teach developers—many of whom are first-time open source contributors—to effectively experiment, learn, and contribute to open source communities. The Cloud Native Computing Foundation (CNCF) uses these programs as a CNCF mentorship platform. TiKV is a CNCF-graduated project. It proposed the Coprocessor support ENUM/SET project.

Introduction to ENUM and SET

When you create a table, two of the values you specify are ENUM and SET. These are MySQL string objects. ENUM defines the permissible values for a table column. SET represents a string object that has zero or more values. When you create a table, you choose the values from a predefined list. In this example, ENUM defines five possible sizes for the shirts table:

CREATE TABLE shirts (
    name VARCHAR(40),
    size ENUM('x-small', 'small', 'medium', 'large', 'x-large')
);

When we create a table, we can specify the table type as ENUM and provide all optional values. Later, when we insert a value into the table, we must use the specified value. Otherwise, an error is reported:

INSERT INTO shirts (name, size) VALUES ('dress shirt','large'), ('t-shirt','medium'),
  ('polo shirt','small');

SELECT name, size FROM shirts WHERE size = 'medium';
+---------+--------+
| name    | size   |
+---------+--------+
| t-shirt | medium |
+---------+--------+

The usage and implementation in TiDB of SET is similar to that of ENUM. We won’t discuss it in detail here. For detailed information about ENUM and SET, you can check the MySQL documents: The ENUM Type and The SET Type.

Why we need this project

In this section, I’ll briefly introduce TiKV Coprocessor and then explain why I built the Coprocessor support ENUM/SET project.

TiDB is an open-source, distributed SQL database, and TiKV is its storage layer. The following figure shows their relationship:

TiDB server is a stateless SQL layer that processes users’ SQL queries, accesses data in the storage layer, and returns the corresponding results to the application. TiKV server is TiDB’s underlying storage engine.

TiDB architecture

We can construct this model to show how TiDB and TiKV interact with each other:

  1. App sends SQL statements to TiDB.
  2. TiDB queries TiKV to get the underlying data.
  3. TiDB performs calculations on the data and returns the results to App.
TiDB and TiKV interaction model

The model has these problems:

  • There is a high transmission overhead between TiDB and TiKV. Take COUNT in the figure as an example. The application only needs to know the result of COUNT, but TiDB still needs to get all the key-value (KV) pairs from TiKV.
  • The load on the TiDB node is heavy, but the load on TiKV is light. Resources are not fully used.

Therefore, engineers proposed that part of a calculation task can be handed over to TiKV, and the TiKV module responsible for this work is called Coprocessor. This work process is also called operator pushdown.

In the past, all calculation tasks involving ENUM or SET could not be pushed down. To solve this problem, I built the Coprocessor support ENUM/SET project.

How I implemented the project

  1. I enabled TiKV to support MySQL ENUM/`SET.

  2. According to Request for Comments (RFC) requirements, I implemented ChunkedVecEnum and ChunkedVecSet.

    ChunkedVec is a new form of memory data expression introduced in TiKV. It represents a column of data. Compared with the row form, it can more effectively use the CPU cache, and it optimized single data, multiple data (SIMD) computing. See the RFC.

    To completely implement ChunkedVecEnum and ChunkedVecSet, we needed to implement decoding. However, it was difficult to estimate the time required to accomplish this, so we couldn’t implement our original plan, which was to have TiDB perform this function. However, I was able to implement these functions in TiKV.

  3. I added Enum and Set evaluation types to VectorValue. Because enum VectorValue was widely used in Coprocessor’s parts, it took me a lot of time to carefully handle the new cases where the enum is used.

  4. I modified codegen-related logic to make it support code with input parameters EnumRef and SetRef. After my pull request (PR) was merged, we could implement the ENUM- and SET-related functions:

    #[rpn_fn]
    #[inline]
    fn cast_enum_as_int(val: EnumRef) -> Result<Option<Int>> {
        Ok(Some(val.value() as Int))
    }
    
    #[rpn_fn]
    #[inline]
    fn cast_set_as_int(val: SetRef) -> Result<Option<Int>> {
        Ok(Some(val.value() as Int))
    }
    

    The most exciting thing is that we can implement these functions without waiting for TiDB to support pushdowns. In order to support ENUM and SET pushdown, both TiDB and TiKV need to implement it. But my work makes it possible to implement the TiKV side without TiDB’s support.

  5. I implemented ENUM and SET support for each aggregate function, including COUNT, FIRST, SUM, AVG, MAX and MIN. This way, I made TiKV support ENUM and SET.

Lesson learned during community cooperation

Export the mod before compilation

When I was trying to implement ChunkedVecSet, I had a disagreement with a repository Maintainer. Because I modified enums, to avoid introducing too many changes in a single PR, I implemented the mod but didn’t export it. As a result, the mod couldn’t be compiled.

I planned to complete this part in my future PR, but the Maintainer insisted that a PR should be the smallest compilable unit.

Honestly, I think he’s right. If everyone does things like me, TiKV’s master branch will get more and more non-compilable code. After many iterations, this PR finally got the Maintainer’s approval.

In hindsight, exporting the mod was essential. After we exported it, we found issues in the mode exporting implementation. But we didn’t notice them, because we hadn’t compiled it before. In addition, when I fixed compilation problems, I got a better understanding of the code involved in this part. This made my follow-up work smoother.

Add a macro expansion test when implementing a macro

After I implemented ChunkedVec{Enum,Set}, I implemented a macro that supported Enum and Set parameters.

A similar story occurred again. The Maintainer asked me to add a macro expansion test. I wanted to add it in my next PR, but the Maintainer thought that the test was highly related to this PR, so that it should be included in this PR.

Therefore, I decided to add a macro expansion test in the PR. To avoid the impact of whitespace characters on the content, my initial implementation was to remove all whitespace and line breaks. However, this changed code semantics. Because I’m a perfectionist, I implemented an algorithm based on TokenTree comparison. As long as all valid tokens were the same, the semantics for the two pieces of code are the same. Finally, this algorithm was applied to the module’s all macro expansion tests.

What’s next

So far, we’ve implemented all the features that are needed for ENUM and SET pushdowns in TiKV. Next, we’ll completely implement decoding and add this support in TiDB. This way, users can smoothly push down ENUM- and SET-related functions. This will make their calculations faster.

For this purpose, I proposed a PR: copr: Roadmap to chunk-based Enum/Set support in TiKV. I’m working on it with TiDB developers. Maybe this year the PRs will be merged.

Thanks to my mentor @skyzh for his guidance. His advice enabled me to complete this project. I’d also like to thank @andylokandy and @zhongzc and developers on the sig-copr Slack channel. Their reviews greatly improved my understanding of Rust and helped me understand what makes good code. This may be my biggest gain in this project.