How we built our Real-time Analytics Platform

http://blog.maxcdn.com/learned-stop-worrying-love-logs/

by jdorfman • 12 years ago

17 11 12 years ago

blog.maxcdn.com

dueprocess• 12 years ago

I stopped using MaxCDN since a website I took offline months earlier suddenly achieved "bandwidth overage" according to them, and my credit card was charged.

Recently, they advertised they now provide server logs to prove usage, but I'm done with them.

mxpxrocks102• 12 years ago

hi hi - Sorry that happened. If you have a sec can you drop me details? chris at maxcdn com

j_s• 12 years ago

TL;DR: TokuMX (MongoDB upgrade) through Redis with Go on the back end and Node.js for the API.

The biggest take-away for me was that someone is selling ($2500/server/year for hot backup + support) an improved ('accelerated speed for inserting data as well as compression') GPL2 MongoDB.

mxpxrocks102• 12 years ago

nice.

nodesocket• 12 years ago

Awesome job. Curious did you investigate fluentd (http://www.fluentd.org)?

Also, seems a bit odd to use MongoDB for write optimized work loads, since it essentially has a global write lock. Doesn't Cassandra perform better at writes?

opendais• 12 years ago

https://blog.serverdensity.com/does-everyone-hate-mongodb/

They use it as well for similar purposes. So it isn't a strange choice.

That said, I'd use Cassandra as well. ;) I found when testing the write performance was significantly better. However, I didn't test the special Mongo variant they are using.

kiyoto2• 12 years ago

Fluentd maintainer here. A lot of folks indeed use Fluentd for application log collection. Perhaps because we embrace JSON as the core data structure, MongoDB is a common data destination (Second most popular plugin) http://www.fluentd.org/plugins/all

bconklin422• 12 years ago

We did prototype for a while with fluentd as well as Cassandra and a few other technologies which we decided not to use in the end. The answer to why we chose a lot of what we did was to satisfy our requirements. The "Minimal Hardware with Redundancy And Scale" and "Minimal Software Layers" requirements came into play in regards to your question. We found that most off the shelf technologies are great at scaling, but they have a cost that would have required us to employ a lot of hardware. We were able to reach the speed we needed by writing our own Go client to ship off the logs to our cluster, and that also gave us the customizability and control to be able to make sure we weren't consuming all of the resources our servers needed during peak times to deliver content.

As for why we chose the database we did, while TokuMX is basically a drop in replacement for MongoDB, it has features that make it much more usable for our situation. Specifically, a compression rate of over 80% without negatively impacting our insert speed and document level locking. And because TokuMX works as a drop in replacement for MongoDB, we were able to use the MongoDB driver (mgo http://labix.org/v2/mgo) for Go which is just a really great tool. Additionally, we are now able to leverage the MongoDB aggregation framework which allows us to use the data in a way that builds very helpful aggregations of the data very quickly.

kiyoto2• 12 years ago

We were able to reach the speed we needed by writing our own Go client to ship off the logs to our cluster

That's a fair argument: when written intelligently, Go has great performance.

The other reason is performance v. flexibility. Any generic log collector (Logstash, Flunetd, etc.) needs to support multiple data sources and outputs and is probably less optimized for the specific situation at hand.

Just curious: what's the buffering strategy for your custom agent written in Go? Does it support file-/memory-based buffering?

cheald• 12 years ago

One of the draws of TokuMX as a replacement for Mongo is that it has document-level locking rather than collection-level locking.

How we built our Real-time Analytics Platform

Related Stories