vreplication: file:pos flavor #5432
Conversation
Signed-off-by: Sugu Sougoumarane <[email protected]>
Add Flavor as a conn param. Add code to handle flavor-specific GTID. Signed-off-by: Sugu Sougoumarane <[email protected]>
In this scheme, the filePos reader detects whether we are in a transaction or not, and emits appropriate GTID events. Signed-off-by: Sugu Sougoumarane <[email protected]>
The vstreamer sent GTIDs "as they came". With the new change, GTIDs are sent only when they matter: on COMMIT, DDL or OTHER. This new approach makes the protocol easier to understand. Also, it makes it easier for filePos to continuously send file and position. The correct values will get used when significant events like COMMIT are encountered. Signed-off-by: Sugu Sougoumarane <[email protected]>
The vplayer currently uses ambiguous rules about how it handles the case where a stop position was exceeded. As part of this change, we'll standardize on: A stop position is considered to be successfully reached if the new position is greater than or equal to the specified position. The main motivation for this change is that the possibility of position mismatch is higher in the case of file:pos tracking. We're likely to hit many false positives if we're too strict. Signed-off-by: Sugu Sougoumarane <[email protected]>
Signed-off-by: Sugu Sougoumarane <[email protected]>
Signed-off-by: Sugu Sougoumarane <[email protected]>
There was a problem hiding this comment.
@sougou - This makes sense to me and the implementation is more straight forward than I was thinking :D
I think we can add more tests later.
The only part I'm a bit hesitant is the changes in the logic to stop position. If you could elaborate a bit more on why that is safe or why it won't introduce any regressions, I think we should merge this.
I'm going to approve but let you merge after talking about the stop position bit.
|
@sougou and I discussed about the stop changes in behavior. The main gist is:
I'm going to merge this and integrate it in #5289 |
Fixes #5424
Details of the implementation are as described in the issue.
The tests are a little light right now. We'll probably need to add more as the flavor evolves.