Locking mechanism to safely reload or restart service

[nhuray]

Context

We are using Consul Template to manage dynamically the configuration of our apps.

Depending of the application we might have 3 behaviours:

• The configuration is reloaded automatically by the process when it change (e.g watch in PM2)
• The process MUST be reloaded running a command (e.g nginx -s reload)
• The process MUST be restarted to take the changes in consideration

For the 2 last cases, we want to reload / restart services one by one in order to prevent service interruption.

See this issue: https://github.com/hashicorp/consul-template/issues/340

Implementation

Consul already provides a lock mechanism to create a distributed semaphore based on Consul Key: https://www.consul.io/docs/commands/lock.html

We want to implement this mechanism wrapping the complexity to configure consul-template and consul lock command in a script ct-supervise :

ct-supervise -n 1  'nginx -s reload'

Options

These options might be passed to the ct-supervise script

• monitor-retry - Retry up to this number of times if Consul returns a 500 error while monitoring the lock. This allows riding out brief periods of unavailability without causing leader elections, but increases the amount of time required to detect a lost lock in some cases. Defaults to 3, with a 1s wait between retries. Set to 0 to disable.

--n: Optional, limit of lock holders. Defaults to 1. The underlying implementation switches from a lock to a semaphore when increased past one. All locks on the same prefix must use the same value.

--name: Optional name to associate with the underlying session. If not provided, one is generated based on the child command.

--pass-stdin: Pass stdin to child process.

• -try: Attempt to acquire the lock up to the given timeout. The timeout is a positive decimal number, with unit suffix, such as "500ms". Valid time units are "ns", "us" (or "µs"), "ms", "s", "m", "h".

• -verbose: Enables verbose output.

The only required argument is the command to run:

ct-supervise  'nginx -s reload'

[ritxos]

Hi @nhuray,

anything on this, I'm curiously looking out for rolling restarts?

[nhuray]

Hi @ritxos,

I didn't start working on this I just discussed it long time ago: https://github.com/hashicorp/consul-template/issues/340

We are using a workaround incrementing the splay setting in Consul Template to be sure the processes are not restarting at the same time. For sure it's not really deterministic and a locking strategy for rolling restarts is better. I might have time to work on it after Xmas.