Gather follows sources of line-based output. It can tail processes or files.
It's controlled via a UNIX domain socket placed at /tmp/gather.
The API is simple. It has two commands: add and rm. Output is printed to
stdout, prefixed by the given source id. Application output and errors are
logged to stderr, separate from source output.
Lines are truncated at 4k bytes.
Example:
# Send commands to gather with socat via the socket
echo 'add file syslog /var/log/syslog' | socat - UNIX-CONNECT:/tmp/gather
echo 'add proc hello echo hello from a process' | socat - UNIX-CONNECT:/tmp/gather
echo 'rm syslog' | socat - UNIX-CONNECT:/tmp/gather
File sources are added to an inotify watch list. The files are not tracked by
name but by inode, meaning rename(2) and unlink(2) do not affect an
attached source. That also means it's not possible to track a file through
rotation. Truncation logic is best effort: if a writer does truncate(2)
followed by a quick write, before the first inotify event can be read, Linux
coalesces the two IN_MODIFY events and the fact that the file was truncated
may be lost depending on the number of bytes that result in the file. The
inotify fd is polled with epoll(7) along with an eventfd(2) that is used as
a side channel to interrupt the blocking read on the epoll fd.
Processes are spawned with Go's os/exec library with stdout piped back to the
parent.
Inotify, epoll, eventfd. This pretty much makes gather Linux-only.
Other sources of output are conceivable, like sockets. An ls command would be
nice to list all currently attached sources. File sources are not tracked
through renames, as they are followed by inode.