[logs] Reliability of SELP (was syslog/tcp ...)

Previous message: Mikael Olsson: "[logs] TCP delivery reliability (was: re: syslog/tcp (selp))"
In reply to: Jason Royes: "Re: [logs] RE: syslog/tcp (selp)"
Next in thread: Jason Royes: "Re: [logs] RE: syslog/tcp (selp)"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

betat_private

I think the picture we've got is getting quite clear.

Neither Syslog Classic (RFC 3164) nor SELP is syslog-reliable (RFC
3195). RFC 3194, via UDP, is utterly unreliable, gracefully handling
overloads by losing as much as needed to avoid getting overloaded.

SELP doesn't guarantee perfectly reliable delivery in the case of
errors in transit or on the server (due to TCP buffering and lack of
precise reporting of successful/failed delivery of each bit of data
back up to the app), but in normal operation it's substantially
reliable, slowing down senders to the rate that the receiver can
cope with rather than losing traffic.

Reliable delivery requires a handshake at the application layer,
where the receiver officially notifies the sender that it has
accepted delivery for each message, allowing the sender to forget
about it. That's where we turn to RFC 3195.

-Bennett