capnslog: why check ppid = 1 for systemd?

[philips]

Current algorithm for journal detection:

1. Check for ppid == 1 https://github.com/coreos/pkg/blob/master/capnslog/init.go#L41
2. Check for an actual journal https://github.com/coreos/pkg/blob/master/capnslog/journald_formatter.go#L28

This is wrong for two reasons: 1. systemd isn't always pid 1 (user session) 2. it is code stutter having a heuristic and then a authoritative check after 3. even if systemd is pid, your parent might not be pid 1.

@barakmich Was there a reason for doing this check that I might be missing? Otherwise we should fix it up.

Thanks.

[barakmich]

This aimed for the 80% case. Open for fix up suggestions. I take it you'd
like to always do journald if available. I kind of hate that, as running on
a system that happens to have journald means you get no log output in the
run-from-the-command-line case.

Is there a better detection algorithm? If so, let's do it (detect parent
name?). Or write to two outputs (stderr and journald) if detected? Many
options, not sure which is preferable.
On Sep 28, 2015 11:23 PM, "Brandon Philips" notifications@github.com
wrote:

Current algorithm for journal detection:

1. Check for ppid == 1
   https://github.com/coreos/pkg/blob/master/capnslog/init.go#L41
2. Check for an actual journal
   https://github.com/coreos/pkg/blob/master/capnslog/journald_formatter.go#L28

This is wrong for two reasons: 1. systemd isn't always pid 1 (user
session) 2. it is code stutter having a heuristic and then a authoritative
check after 3. even if systemd is pid, your parent might not be pid 1.

@barakmich https://github.com/barakmich Was there a reason for doing
this check that I might be missing? Otherwise we should fix it up.

Thanks.

—
Reply to this email directly or view it on GitHub
#38.

[jlamillan]

What about util.RunningFromSystemService() from coreos/go-systemd?

[xiang90]

@jonboulle How do you think we should deal with this? Sending to journal by detecting pid+systemd is proven to confuse people, in rkt case at least:

etcd-io/etcd#5449
etcd-io/etcd#5779

[jonboulle]

@xiang90 for this specific issue I'm sympathetic towards #38 (comment)

I'm actually not convinced that capnslog should log directly to the journal - rather it seems to me it'd be sufficient if it didn't log timestamps when it detects it's invoked from systemd.

[xiang90]

rather it seems to me it'd be sufficient if it didn't log timestamps when it detects it's invoked from systemd.

If the timestamp is the only issue, yea, probably we should just try #38 (comment) to detect systemd, and do not log timestamp.