TiFlash panics with Too many open files in the cloud GCP env

[solotzg]

Bug Report

Please answer these questions before submitting your issue. Thanks!

1. Minimal reproduce step (Required)

• Topology: 3-tidb-16C32G / 9-tikv-16C64G-2000G / 3-pd-4C15G-50G / 2-tiflash-16C128G-500G
• Platform: GCP
• Unknown workloads

2. What did you expect to see? (Required)

• No panic

3. What did you see instead (Required)

Too many fd under tiflash process. The number of FD keeps growing, causing queries to fail and eventually tiflash to panic. Most of the FDs are related to sockets and a large amount of sockets are still open but can not be found in /proc/net.

sh-5.1# ls -l /proc/1/fd/ | wc -l
295268
sh-5.1# ls -l /proc/1/fd/ | grep "eventfd" | wc -l
98368
sh-5.1# ls -l /proc/1/fd/ | grep "eventpoll" | wc -l
98391
sh-5.1# ls -l /proc/1/fd/ | grep "socket" | wc -l
98492

Other

In the AWS environment, there is no such problem yet.

4. What is your TiFlash version? (Required)

v7.5.3

[solotzg]

After disabling mpp, the number of socket fd no longer continues to grow. There may be potential bugs in the implementation about mpp.

set global tidb_allow_fallback_to_tikv = "tiflash";
set global tidb_allow_mpp = 0;
set global tidb_allow_tiflash_cop = 1;