[client-v2] ZSTD is the default but the zstd-jni "cloud" classifier ships no Windows or Intel-macOS natives, so every query fails there

未关闭
#3,175 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
3/5
预计耗时
1-2 天
新手友好度
72/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
java
领域
backend, databases

调研方向

Start with client-v2/pom.xml at lines 61–66 and the parent dependency management for zstd-jni; compare the classifier and version with the other modules. Verify the resolved artifact includes win/amd64 and darwin/x86_64; the issue suggests a unit test for win/amd64 if cross-platform CI is unavailable. Done means default-compression queries work on the affected platforms without missing-native errors.

由索引模型根据 Issue 内容生成。

描述

Description

client-v2 depends on zstd-jni with the cloud classifier
(client-v2/pom.xml:61-66):

<dependency>
    <groupId>com.github.luben</groupId>
    <artifactId>zstd-jni</artifactId>
    <version>1.5.7-20</version>
    <classifier>cloud</classifier>
</dependency>

The cloud classifier jar contains native libraries for only three platforms:

darwin/aarch64/libzstd-jni-1.5.7-20.dylib
linux/aarch64/libzstd-jni-1.5.7-20.so
linux/amd64/libzstd-jni-1.5.7-20.so

There is no win/* and no darwin/x86_64. The default zstd-jni artifact of the same
version contains 18 natives, including win/amd64, win/x86, win/aarch64 and
darwin/x86_64.

Because ClientConfigProperties.COMPRESSION_METHOD defaults to CompressionMethod.ZSTD
(ClientConfigProperties.java:104),
ZSTD is not an opt-in path. On Windows and on Intel macOS, client-v2 and jdbc-v2 fail on
the first query with default settings — the class initializer for
com.github.luben.zstd.Zstd throws, and every subsequent call fails with
NoClassDefFoundError.

Steps to reproduce

On Windows x86-64, against a server new enough to return ZSTD-compressed blocks (26.x):

try (Connection conn = DriverManager.getConnection("jdbc:clickhouse://localhost:8123/", "default", "");
     Statement stmt = conn.createStatement();
     ResultSet rs = stmt.executeQuery("SELECT 1")) {
    rs.next();
}

No compression setting is applied — this is the default configuration.

Equivalently, mvn -pl jdbc-v2 verify -Dit.test=DatabaseMetaDataTest on Windows against the
latest image gives 44 of 47 failures, all from this one cause. The same command with
-DclickhouseVersion=25.8 passes 47/47, because that server does not send ZSTD blocks.

Error Log or Exception StackTrace

First failure — the class initializer:

java.lang.UnsatisfiedLinkError:
Failed to open resource /win/amd64/libzstd-jni-1.5.7-20.dll as stream:
Unsupported OS/arch, cannot find /win/amd64/libzstd-jni-1.5.7-20.dll or load zstd-jni-1.5.7-20
from system libraries. Please try building from source the jar or providing
libzstd-jni-1.5.7-20 in your system.
	at com.github.luben.zstd.util.Native.load(Native.java:124)
	at com.github.luben.zstd.util.Native.load(Native.java:93)
	at com.github.luben.zstd.Zstd.<clinit>(Zstd.java:22)
	at com.clickhouse.client.api.internal.CompressedBlockInputStream.refill(CompressedBlockInputStream.java:156)
	at com.clickhouse.client.api.internal.CompressedBlockInputStream.read(CompressedBlockInputStream.java:63)
	at com.clickhouse.client.api.data_formats.internal.BinaryStreamReader.readByteOrEOF(BinaryStreamReader.java:1555)
	at com.clickhouse.client.api.data_formats.RowBinaryWithNamesAndTypesFormatReader.readSchema(RowBinaryWithNamesAndTypesFormatReader.java:41)
	at com.clickhouse.client.api.Client.newBinaryFormatReader(Client.java:2451)
	at com.clickhouse.jdbc.StatementImpl.executeQueryImpl(StatementImpl.java:325)

Every later query in the same JVM:

java.lang.NoClassDefFoundError: Could not initialize class com.github.luben.zstd.Zstd
	at com.clickhouse.client.api.internal.CompressedBlockInputStream.refill(CompressedBlockInputStream.java:156)
	...
Expected Behaviour

The default configuration should work on every platform the driver supports. A JDBC driver
that cannot run SELECT 1 on Windows out of the box is unusable there, and the failure
surfaces as NoClassDefFoundError rather than anything a user can act on.

Root cause

Introduced in 9dc6a95 ("Added ZSTD block compression support"), merged as #3157, which
fixed #3105 and #3120. Those fixes are correct; only the choice of classifier is the
problem here. The
cloud classifier is published by zstd-jni for deployments that only ever run on Linux or
Apple-silicon containers; it is a size optimisation (1.0 MB vs 6.5 MB) and is not suitable
for a general-purpose client library.

Note that clickhouse-data, clickhouse-http-client and clickhouse-jdbc already depend
on the default zstd-jni artifact through dependencyManagement
(${zstd-jni.version}, currently 1.5.5-5, no classifier), so client-v2 is the only module
using the restricted jar, and it also pins its own version instead of the managed one.

Suggested fix

Preferred — drop the classifier and use the managed version, so the module matches the rest
of the project:

<dependency>
    <groupId>com.github.luben</groupId>
    <artifactId>zstd-jni</artifactId>
</dependency>

with <zstd-jni.version> in the parent raised to 1.5.7-20. Cost is roughly 5.5 MB of jar
size; correctness on Windows and Intel macOS seems worth that, and users who need the small
jar can still exclude it and substitute the cloud artifact themselves.

Independently of the dependency choice, CompressedBlockInputStream could fail more
usefully: catching UnsatisfiedLinkError / NoClassDefFoundError when ZSTD is first
touched and reporting that the native library is missing for this platform — naming the
compression.method setting — would turn an opaque NoClassDefFoundError into something
actionable, and would protect against the same class of problem in future.

A regression test would need a non-Linux runner; the CI matrix is Linux-only, which is why
this was not caught. At minimum, a unit test asserting that the resolved zstd-jni artifact
contains win/amd64 would catch a reintroduction without needing a Windows runner.

Configuration
Environment
  • Cloud
  • Client version: 0.11.0-rc1-SNAPSHOT (main at 16dcde46)
  • Language version: JDK 21 (compiled with release 17)
  • OS: Windows 11 x86-64
ClickHouse Server
  • ClickHouse Server version: clickhouse/clickhouse-server:latest via Testcontainers
  • ClickHouse Server non-default settings, if any: none
主要语言
Java
星标
1.6k
派生
649
平均合并
1 天 20 小时
30 天内合并 PR
39

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

ClickHouse/clickhouse-java 的其他 Issue

查看 ClickHouse/clickhouse-java 的全部 Issue

相似的 Issue

更多 Java Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。