19:00 Harrison says that it takes 18 secs to determine the master has failed. 19:25 - he says only couple events are missing. Why is that? I would expect a lot of events missing after 18 secs of master failure. Do we still write into Log tails while master is down?
For master to fail, it has to be true that all log tailers and slaves not being table to talk to it. If there's one of them being able to talk to it, the master has not fail. There's just a network partition. So, there are two scenarios: 1. no one can talk to master - in this case, master **cannot accept writes**. Hence, nothing missing. 2. one or more can talk to master - in this case, master has not failed. Hence, nothing missing. What he meant by a couple missing is actually amongst slave & log tailers themselves. Since master send binlogs to all of them, there could be some package loss or delay that cause some of them being able to receive earlier than others.
I am assuming how mysql scales to support millions of users in facebook messenger is by there are multiple mysql instances which are deployed per region and there is router upstream to store the messages for users from a particular region to that particular mysql instance , is my understanding correct ?
Thank you for saying My S-Q-L and not my "sequel"
Explained very nicely :) Amazing stuff !!!
Whats the use of having multiple log tail-ers per machine ?
isn't it the same as adding very active write-back cache in front of mysql?
very good presentation.
19:00 Harrison says that it takes 18 secs to determine the master has failed.
19:25 - he says only couple events are missing.
Why is that? I would expect a lot of events missing after 18 secs of master failure.
Do we still write into Log tails while master is down?
For master to fail, it has to be true that all log tailers and slaves not being table to talk to it.
If there's one of them being able to talk to it, the master has not fail. There's just a network partition.
So, there are two scenarios:
1. no one can talk to master - in this case, master **cannot accept writes**. Hence, nothing missing.
2. one or more can talk to master - in this case, master has not failed. Hence, nothing missing.
What he meant by a couple missing is actually amongst slave & log tailers themselves. Since master send binlogs to all of them, there could be some package loss or delay that cause some of them being able to receive earlier than others.
love it !
are the slides published?
I am assuming how mysql scales to support millions of users in facebook messenger is by there are multiple mysql instances which are deployed per region and there is router upstream to store the messages for users from a particular region to that particular mysql instance , is my understanding correct ?
Is the Iris queue *per conversation*?
It looks like it's one queue per user
But, how is the MySQL sharded? Consistent hashing or something else?