Around my last response, I had 210,000,000 records in my table. I continued with my 'slow delete' process, roughly 30-50 rows per second for the last 6 days.
Please note this is entirely unscientific.
I've gone from 210,000,000 to 188,396,007 actual rows.
The most accurate is 5.6. I suspect 5.5.41 could have been more accurate but the problem is that I did not run an analyze on it 6 days ago so difference has been accrued over a much larger time frame. My suspicion is that it would be much more accurate had I done this and that for *my* data churn model, the old style row estimations (toku 7.5.5) were much more accurate.
During this time, no tables had an analyze performed on them, and auto_analyze was off.
At this rate, I expect my PS 5.7.13-6-log to hit a estimated row count of 0 well before the table actually has 0 rows.
I personally would consider this a bug - with auto_analyze being a just barely acceptable workaround as this particular workload is infrequent.
Ok, here's my follow up.
Around my last response, I had 210,000,000 records in my table. I continued with my 'slow delete' process, roughly 30-50 rows per second for the last 6 days.
Please note this is entirely unscientific.
I've gone from 210,000,000 to 188,396,007 actual rows.
Here are my show table status rows:
PS 5.7.13-6-log - 150,815,123 tokudb- 7.5.5-MariaDB- log - 193,416,244
PS 5.6.31-77.0-log - 185,383,998
5.5.41-
The most accurate is 5.6. I suspect 5.5.41 could have been more accurate but the problem is that I did not run an analyze on it 6 days ago so difference has been accrued over a much larger time frame. My suspicion is that it would be much more accurate had I done this and that for *my* data churn model, the old style row estimations (toku 7.5.5) were much more accurate.
During this time, no tables had an analyze performed on them, and auto_analyze was off.
At this rate, I expect my PS 5.7.13-6-log to hit a estimated row count of 0 well before the table actually has 0 rows.
I personally would consider this a bug - with auto_analyze being a just barely acceptable workaround as this particular workload is infrequent.