Github user viirya commented on the pull request:
https://github.com/apache/spark/pull/4467#issuecomment-74010907
Let's analyze it clearly. The following is a simplified status
transformation of the problem:
tracker receivers
t = 1 started registered:{A, B}; starting, not registered: {C}
t = 2 stopping got stop msg:{A, B}; starting, not registered: {C}
t = 3 stopping stopped:{A, B}; registered: {C}
The above causes potential data loss. We want to avoid that. I agree.
If we implement option 1, now the status transformation:
tracker receivers
t = 1 started registered:{A, B}; starting, not registered: {C}
t = 2 stopping got stop msg:{A, B}; starting, not registered: {C}
*we are going to wait for receivers that are started but not registered yet.
*suppose we wait a fixed time period n.
*however, we can't guarantee when the receiver C will be registered.
*so, after waiting time n, the system status can be:
t = n+2 stopping stopped:{A, B}; registered: {C}
*or still
t = n+2 stopping stopped:{A, B}; starting, not registered: {C}
As you see, there will still be possible status that we have unregistered
receiver C that processes data.
This pr implements another approach. The receivers register first then do
starting process:
tracker receivers
t = 1 started registered, started:{A, B}; registered, starting:
{C}
t = 2 stopping got stop msg:{A, B, C}; **D wants to register ->
timeout
t = 3 stopping stopped:{A, B, C}
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
