schumihan

Cassandra is a good fit for write heavy workload, as it is optimized for write...

Agreed. You can achieve very high productivity if you use Scala and Java properly.

I just notice VIA and HTC are founded by the same woman[1], and I know the financial situation of VIA is not good...

[1] http://blog.laptopmag.com/2011-most-important-women-in-mobil...

"A user's data is stored in a single row in HBase"

Some column families of HBase can be accumulated to hundreds of mega bytes very easily, I'm afraid...

The token in cookie stored by browser is portable, either. How can web apps prevent the attackers exploiting the system from stealing the cookie token?

When different customers use different key to encrypt the data, even if the source is the same, the outcome is different. So it is not feasible to do de-duplication across different customers' data, unless all of them use the same key.

Does SpiderOak only provide backup service? Erasure encoding is efficient for cold data. Do you use erasure encoding to distribute the hot data across clusters?

I use patch queue[1] to manage these short lived branch, and it works for my purpose. You just pop the patches and delete them all when you want to prunge the branch. It also make it easy for others to review your code.

If these short-lived branch will be merged back shortly. Bookmark is another alternatives.

[1] http://mercurial.selenic.com/wiki/MqExtension [2] http://mercurial.selenic.com/wiki/BookmarksExtension

If I tries to shorten the url of http://google.com several times, goo.gl always give back different result. I just wonder will goo.gl merge theses url together sometime later?

This interesting found implies how goo.gl is designed. They prefer writing directly to avoid duplication check and random seek caused by the check. And they though most of the urls are cold so it is not essential to merge them together to keep the hot cache small.

Or they assume nobody wants to play the system in this way. :-)

The architecture of MegaStore is so complicated... There are so many roles in the system, e.g. stateless replicate server, different stateful paxos instances, stateful co-coordinator, bigtable and other auxilary roles.

I suspect some roles are added later because the original implementation does not meet their design goal, as patch to normal software product.

It is a nightmare for operation engineers of GAE.

Unlike Bigtable/HBase, cassandra is fully distributed, either, just as Dynamo...

For any SQL/NoSQL system, the more RAM, the better. They need as much RAM as possible to cache the data.

If you observed cassandra consumes a lot of RAM even with small dataset, you should tune the GC settings. Besides it, JVM itself consumes a lot of RAM.

Most of the time, I use both as a tool to do visual diff, instead of code merge.

KDiff3 is not so good as winmerge as visual diff tool. And kdiff3 UI has small problem on Ubuntu, as it was not developed for gnome environment.

+1 for beanstalkd. Zero deployment and blazing fast!

Why are there so many "best papers" on CHI 2010?

I like winmerge. Even kdiff3 is not as useful as winmerge, IMHO.

http://kdiff3.sourceforge.net/

On their hardware appliance, they use a special instruction to implement the barrier. For software appliance, the performance may be a real problem.

I think this is reasonable.

My guess: Gmail wants you to invite somebody in your contact list directly. And they are reluctant to reset the cap of invitation because there are so many users in the system.

But why do they not hide the interface directly?

Can we build a system above ZeroMQ for such purpose easily? I think the async semantic of ZeroMQ is suitable to mitigate the latency of multiple RPC.

It's very difficult to avoid it when thing goes complicating. A may reference B, B may reference C, ... and somewhere, X may reference back to A. If this strategy makes sense for big project, many hackers work with C++ can use it either, as the performance overhead is very low.

But in the real world, reference-count based GC just has limited usage.

Reference count based GC? How would you deal with circular reference?

Is it similar to ZeroMQ(http://www.zeromq.org/)?

Are you use 0.6.x branch or 0.7 branch? 0.6.x is quite stable in my deployment.

I suppose fb caches the recent messages in memory, and use HBASE as archive storage most of the time.

It depends on the developer's favor to pick up a system. If it's me, I will pickup cassandra. The consistency model is not so hard to work with.

UDP is more easier to penetrate a firewall, IMHO...

You can work with much higher productivity in java...

I don't think so. Percolator does not support stream operators, and it supports transaction semantic. Streaming process engine like s4 will never handle the race conditions brought by multiple writers.

I'm one of them. The weakness of yuku is most of the videos are uploaded by the website itself, not the user generated content. The reasons behind it is quite complicated...