[MESOS-8481] Agent reboot during checkpointing may result in empty checkpoints.

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Fix Version/s: 1.5.0
• Target Version/s: 1.5.0
• Sprint: Mesosphere Sprint 73
• Assignee: Michael Park
• Reporter: Chun-Hung Hsiao
• Shepherd: Vinod Kone
• Created: 24/Jan/18 01:11
• Updated: 30/Jan/18 06:36
• Resolved: 30/Jan/18 06:36

Description

An empty checkpoint file was created due to the following incident.

At 17:12:25, the master assigned a task to an agent:

I0123 17:12:25.000000 18618 master.cpp:11457] Adding task 5602 with resources cpus(allocated: *):0.1; mem(allocated: *):128 on agent aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4 at slave(1)@<agent_ip>:5051 (<agent_ip>)
I0123 17:12:25.000000 18618 master.cpp:5017] Launching task 5602 of framework 6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112 (Balloon Framework OOM) at scheduler-fbba22f7-ebbc-4864-8394-0aa558f8ffaa@<ip>:10015 with resources [...] on agent aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4 at slave(1)@<agent_ip>:5051 (<agent_ip>)

Meanwhile, the agent is being rebooted:

$ last reboot
reboot   system boot  3.10.0-693.11.6. Tue Jan 23 17:14 - 00:09  (06:55)

The agent log did not show any information about the task, possibly because there was no fsync before reboot:

I0123 17:12:09.000000 17237 http.cpp:851] Authorizing principal 'dcos_checks_agent' to GET the endpoint '/metrics/snapshot'
-- Reboot --
I0123 17:15:40.000000  2689 logsink.cpp:89] Added FileSink for glog logs to: /var/log/mesos/mesos-agent.log

However, the agent was checkpointing the task before reboot:

$ sudo stat /var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/executors/5602/
  File: ‘/var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/executors/5602/’
  Size: 39        	Blocks: 0          IO Block: 4096   directory
Device: ca40h/51776d	Inode: 67306254    Links: 3
Access: (0755/drwxr-xr-x)  Uid: (    0/    root)   Gid: (    0/    root)
Context: system_u:object_r:unlabeled_t:s0
Access: 2018-01-24 00:23:43.237322609 +0000
Modify: 2018-01-23 17:12:25.751463030 +0000
Change: 2018-01-23 17:12:25.751463030 +0000
 Birth: -

And since there was no fsync before reboot, all checkpoints resulted in empty files:

$ sudo stat /var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/framework.info
  File: ‘/var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/framework.info’
  Size: 0               Blocks: 0          IO Block: 4096   regular empty file
Device: ca40h/51776d    Inode: 33967500    Links: 1
Access: (0600/-rw-------)  Uid: (    0/    root)   Gid: (    0/    root)
Context: system_u:object_r:unlabeled_t:s0
Access: 2018-01-23 17:15:41.485506070 +0000
Modify: 2018-01-23 17:12:25.749463047 +0000
Change: 2018-01-23 17:12:25.749463047 +0000
 Birth: -
$ sudo stat /var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/framework.pid
  File: ‘/var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/framework.pid’
  Size: 0               Blocks: 0          IO Block: 4096   regular empty file
Device: ca40h/51776d    Inode: 33967495    Links: 1
Access: (0600/-rw-------)  Uid: (    0/    root)   Gid: (    0/    root)
Context: system_u:object_r:unlabeled_t:s0
Access: 2018-01-23 23:00:42.190975780 +0000
Modify: 2018-01-23 17:12:25.749463047 +0000
Change: 2018-01-23 17:12:25.749463047 +0000
 Birth: -
$ sudo stat /var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/executors/5602/executor.info
  File: ‘/var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/executors/5602/executor.info’
  Size: 0         	Blocks: 0          IO Block: 4096   regular empty file
Device: ca40h/51776d	Inode: 67306255    Links: 1
Access: (0600/-rw-------)  Uid: (    0/    root)   Gid: (    0/    root)
Context: system_u:object_r:unlabeled_t:s0
Access: 2018-01-23 17:12:25.751463030 +0000
Modify: 2018-01-23 17:12:25.751463030 +0000
Change: 2018-01-23 17:12:25.751463030 +0000
 Birth: -

So were forked.pid and task.info.

As a result, the agent failed to recover after reboot:

E0123 17:15:41.000000  2709 slave.cpp:6800] EXIT with status 1: Failed to perform recovery: Failed to recover framework 6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112: Failed to read framework info from '/var/lib/mesos/slave/meta/slaves/aaf0a62f-a6eb-4c1d-80db-5fdd26fe8008-S4/frameworks/6f9b0688-38f7-4b38-bb1c-421f55e486e5-0112/framework.info': Found an empty file

The error came from https://github.com/apache/mesos/blob/a892a2e80255291e6cd5cb3b0e90b9a029989c99/3rdparty/stout/include/stout/protobuf.hpp#L367.
This is related to a recent change of the checkpoint recovery logic: we used to ignore empty checkpoint files before, but now it return errors on empty checkpoint files.