Kelly Sommers

Cassandra was built to be crash only for this reason https://lists.apache.org/thread.html/%3CCALamADLmmgmdA2vKYQAXDRf+ngU9YNAK+mVDbSDKmcsnrj3bMQ@mail.gmail.com%3E …

but not anymore it seems?

@kellabyte what’s the redis problem?

@fields Killing a node could lose data if I recall. I forget the details.

@kellabyte Our problem with kill -9 was that the kernel closes TCP connections so neatly. Process death and machine death look different.

there’s remote power API’s or even devices.

hah yeah - tedious to set up and maintain. Nobody does it. kill -9 is a much cheaper starting point.

Would terminating the VM process a tested process was run in simulate machine failure better?

with respect to filesystem, ports, etc, that is

I expect that’s way better, yes. Still different behavior than powering off a disk.

what does a powered-off disk look like to the system call?

@kellabyte Basho used to run parties called "killdashnine" after MongoDB said you weren't supposed to do that to a database.

@aphyr @kellabyte In addition to kill -9, I also like to suspend processes (^z) and resume; simulates full GC for remote java processes

@patrickperalta @aphyr @kellabyte That's a personal testing favorite of mine. That using TCP clients that send/receive max 1 byte/sec.

@patrickperalta @aphyr @kellabyte Ah, typo detection so late, sorry. "Also, using TCP clients that send/receive only 1 byte/sec."

@kellabyte @aphyr power pulls and real network partitions are far more important. So much sneaks through kill -9 :)