3 ms·

We're just discussing over in this thread: http://news.ycombinator.com/item?id=296656 http://news.ycombinator.com/item?id=296656 I have had really good experie

by eries 18y ago

We're just discussing over in this thread: http://news.ycombinator.com/item?id=296656 http://news.ycombinator.com/item?id=296656 I have had really good experience with a very simple centralized lookup facility. It's very easy to prevent that from becoming a CPOF, either using replication or, if you insist, some form of hash-based partitioning of the directory itself. I'm a fan of doing the partitioning at the app-level, and doing it horizontally rather than vertically. The loss of joins is an issue, but in some cases you can vertically partition the parts of the app that need those joins. For the others, I have had good luck using a transactional message-passing system, where one shard can reliably send a message to another shard.

furiouslol 18y ago
Yes. I'm reading your slides now. Very interesting set of technologies you have there. How do you integrate the usage of multiple languages? Do you use Thrift? Instead of a centralized lookup facility, have your team tried a decentralized approach instead where each DB server (node) contains a portion of the lookup directory?
- eries 18y ago
  We mostly used simpler RPC mechanisms like XMLRPC or REST. But cross-language calls (other than client-server) were really the exception. Probably the only really performance sensitive ones were to our SOLR-based search. Yes, you can easily move chunks of the address space around, but we decided that the pain that results from remapping (like when you add a new machine) is worse. We very rarely had a problem with our master (which we ran in master-master replication, how's that for term overloading). The few exceptions were times that our own bad code ran amok driving load. It was way better to invest in an automatic query killer.
  - furiouslol 18y ago
    I'm very interested in knowing how you guys grew so big so fast under the radar without much evangelizing from the SV scene. Hope to read about this in your next blog post!
    - eries 18y ago
      Just for you: http://startuplessonslearned.blogspot.com/2008/09/sem-on-five-dollars-day.html http://startuplessonslearned.blogspot.com/2008/09/sem-on-fiv... Would love your feedback, if this is helpful.
gtani 18y ago
this one too http://news.ycombinator.com/item?id=228581 http://news.ycombinator.com/item?id=228581

by eries 18y ago

furiouslol 18y ago

Yes. I'm reading your slides now. Very interesting set of technologies you have there. How do you integrate the usage of multiple languages? Do you use Thrift? Instead of a centralized lookup facility, have your team tried a decentralized approach instead where each DB server (node) contains a portion of the lookup directory?

eries 18y ago
We mostly used simpler RPC mechanisms like XMLRPC or REST. But cross-language calls (other than client-server) were really the exception. Probably the only really performance sensitive ones were to our SOLR-based search. Yes, you can easily move chunks of the address space around, but we decided that the pain that results from remapping (like when you add a new machine) is worse. We very rarely had a problem with our master (which we ran in master-master replication, how's that for term overloading). The few exceptions were times that our own bad code ran amok driving load. It was way better to invest in an automatic query killer.
- furiouslol 18y ago
  I'm very interested in knowing how you guys grew so big so fast under the radar without much evangelizing from the SV scene. Hope to read about this in your next blog post!
  - eries 18y ago
    Just for you: http://startuplessonslearned.blogspot.com/2008/09/sem-on-five-dollars-day.html http://startuplessonslearned.blogspot.com/2008/09/sem-on-fiv... Would love your feedback, if this is helpful.

gtani 18y ago

this one too http://news.ycombinator.com/item?id=228581 http://news.ycombinator.com/item?id=228581