Crash: 'invalid opcode' in lancedb.linux-x64-gnu.node on Ubuntu 24.04 (glibc 2.39)

[donsonlau898]

环境

项目	版本
OS	Ubuntu 24.04.4 LTS
Kernel	6.8.0-106-generic
glibc	2.39-0ubuntu8.7
Node.js	v22.22.1
npm	10.9.4
OpenClaw	2026.3.24
memory-lancedb-pro	v1.1.0-beta.10
Diagnostic build	memory-lancedb-pro-diag-20260308-0058
Embedding model	jina-embeddings-v5-text-small

问题描述

插件导致 OpenClaw Gateway 持续崩溃重启。插件注册成功后会正常运行约 3-5 秒，然后崩溃。每 28-32 秒崩溃一次，持续约 14 分钟，共崩溃 42 次。

无 JavaScript 错误日志 - 崩溃发生在 native 代码级别，JavaScript 的 try/catch 无法捕获。

错误日志

Kernel 日志显示 lancedb.linux-x64-gnu.node 触发 CPU "invalid opcode" 陷阱：

Mar 26 18:48:26 donsonclaw kernel: traps: tokio-runtime-w[1622069] trap invalid opcode ip:739345583700 sp:73933bdfcf18 error:0 in lancedb.linux-x64-gnu.node[739340e0c000+5dd4000]
Mar 26 18:48:54 donsonclaw kernel: traps: tokio-runtime-w[1622745] trap invalid opcode ip:715591583700 sp:7158297e9f18 error:0 in lancedb.linux-x64-gnu.node[71558ce0c000+5dd4000]
Mar 26 18:49:24 donsonclaw kernel: traps: tokio-runtime-w[1623472] trap invalid opcode ip:7f1b41583700 sp:7f1dcb9faf18 error:0 in lancedb.linux-x64-gnu.node[7f1b3ce0c000+5dd4000]
... (每 28-32 秒重复)

崩溃频率

约每 30 秒一次，共 42 次崩溃（18:48-19:03）

临时解决

删除插件配置后 Gateway 立即稳定。

初步分析

问题似乎是 glibc 2.39 与 lancedb.linux-x64-gnu.node 的兼容性问题，可能与 lancedb#1884 相关。

配置

{
  "plugins": {
    "entries": {
      "memory": {
        "id": "memory-lancedb-pro",
        "config": {
          "embedding": {
            "apiKey": "xxx",
            "model": "jina-embeddings-v5-text-small"
          },
          "llm": {
            "model": "openai/gpt-oss-120b"
          },
          "smartExtraction": true
        }
      }
    },
    "allow": ["memory-lancedb-pro"],
    "slots": {
      "memory": "memory-lancedb-pro"
    }
  }
}

[AliceLJY]

Thanks for the detailed report — the kernel logs and crash frequency data are very helpful.

This is a LanceDB upstream issue: the prebuilt lancedb.linux-x64-gnu.node binary includes AVX2/AVX-512 instructions that your CPU may not support. The invalid opcode trap confirms the CPU is hitting an instruction it doesn't recognize.

This is the same class of issue as #172 (Illegal Instruction on older non-AVX2 Intel CPUs).

To confirm, run:

grep -o 'avx2' /proc/cpuinfo | head -1

If this returns nothing, your CPU doesn't support AVX2 — which is the root cause.

Upstream tracking: this is related to lancedb#1884. Until LanceDB ships a non-AVX2 build or a fallback path, there's no workaround on our side unfortunately.

We'll keep this open and link to #172 for tracking.

[AliceLJY]

Upstream limitation — LanceDB native binary requires AVX2 support. Tracked upstream at lancedb#1884. Nothing we can fix on the plugin side. If LanceDB ships a non-AVX2 build in the future, feel free to reopen.