Skip to content

Latest commit

 

History

History
133 lines (99 loc) · 6.09 KB

README.md

File metadata and controls

133 lines (99 loc) · 6.09 KB
Raw

Why the name?

Greg is for "Global Registrator". It is a scalable logger with a high-precision global time axis.

What is it for?

  • Who hasn't merged logs from different machines and hasn't cursed their out-of-sync clocks and the offline mode of processing?

  • Who hasn't wondered, how much time it takes for a request to travel from one machine to the other, and hasn't cursed his logs for not being able to provide this information?

  • Who hasn't wondered, how on earth one can debug bottlenecks and find critical paths in latency-sensitive distributed systems?

Greg solves these problems by being a global logger with a high-precision global time axis.

It was written precisely because of the desperate need for such a tool and lack of an existing one.

More precisely - I've had the idea of such a tool for a long time, then I wrote an implementation for my employer, I liked it and decided to implement the same idea in open source.

What can I do with it?

You can see that your http request arrived to the server 12ms after it departed from the client, even if the client's and server's clocks are off by 3 hours. Just log "Request departed" at the client and "Request arrived" at the server - you'll see their timestamps will differ by 12ms.

Or you can see that, when you were round-robining requests to 1000 machines through a message queue, the last machine received the request 1.2 seconds after the first one, so your message queue handles 830 messages/sec.

Greg is especially useful in conjunction with the timeplotters tools.

How do I use it?

Download and unpack the distribution from here (haskell binding is available separately on hackage - install it as cabal install greg-client).

Launch the server.

./greg-server.sh (or greg-server.bat)

Use the client API from any machine (a single function).

For example, in Java (bindings for several other languages coming soon, or you can easily write one yourself - see [Protocol]):

Greg.log("Request " + requestId + " departed");

Or use the Java log4j appender (org.greg.client.log4j.GregAppender).

The server will output records to stdout (redirect it somewhere yourself) in the following format (spaces are spaces, not tabs):

MACHINE CLIENT_ID GLOBAL_TIMESTAMP MESSAGE

for example:

DC03-016 SearchService 2010-11-20 14:06:35.782 Query arrived: rick astley
  • MACHINE is where the record was generated
  • CLIENT_ID is a user-specified identifier of the component that generated the record
  • GLOBAL_TIMESTAMP is global timestamp of the record, in the server's time zone
  • MESSAGE

The server will do its best so that messages are logged in order of GLOBAL_TIMESTAMP.

See ConfigurationGuide for some options, in case you need them (you probably don't, except for the option configuring server address and the CLIENT_ID).

You can restart the server at any time - clients will survive it (in the worst case, some up to 1 second's worth of log messages from the client will be lost).

What if I like it?

Then please:

  • Use it and tell me how it goes.
  • Write a client in your favourite language. It's easy! Shouldn't take more than a few hours - see [Protocol].

What if I don't like it?

Then tell me how it could be made better.

How does it work?

Greg is a client-server logger: you've got a single server and clients (bindings for several languages) sending records to it. But do not fear: it's NOT like a remote syslog.

  • Greg registers events at the time of their occurence, not of arrival to the server;
  • Greg computes the clock offset between server and each client, thus aligning all records on a common global time axis with millisecond (or better) precision;
  • Greg uses a high-precision system clock;
  • Greg sends records in large batches and therefore it can handle a very high throughput of them (peak ~400,000 hello-world records per second - about 20MB/s) without sacrificing timing precision;
  • Greg was designed with robustness in mind and, if something fails, the whole thing doesn't crash; it is also designed to not crash under any kind of load (at worst, it will drop some messages).
  • Greg can handle thousands of clients.

Note that this Java implementation has not been tested as extensively as the original one - but some testing has been done, and it works basically in the same fashion. Please try it and tell me if something's broken!

Small script

The significant assumption is that clients and server of Greg must be on a reasonably speedy LAN, where network latencies are almost symmetric (it won't work well if it takes 1s for a packet to travel in one direction and 300ms in the other - unfortunately, it is theoretically impossible to work well under such circumstances, nor is it possible to detect them).

Why not X?

  • Why not ntpd + regular logging?
    • You'd still need to somehow merge logs in off-line after collecting them on ntpd-synced machines, you can't "tail -f" the joint log.
    • Not all systems have ntpd installed.
    • Not on all systems, even where ntpd is installed, you actually have access to a system call which provides the current time with sufficient precision (for example, on Windows you don't, and on some Unixes you don't).
    • When using "precise system time", you're amenable to leaps of logged time at clock changes (like summer/winter time, or leap seconds) - on the contrary, Greg uses a continuous monotonic clock (an interval timer).
  • Why not syslog?
    • You'll still need ntpd for a global time axis - see notes about ntpd.
    • See note about high-precision timer for "regular logging".
    • syslog won't output the result in ascending order of global time.
    • Try piping 400,000 or even 50,000 messages per second into syslog.
  • Why not Facebook's scribe?
    • It does not provide a global time axis.