Whenever we use fallback on a table, there are two copies of the table:
one is Primary copy and other one is Fallback copy.
If an amp failure occurs when using fallback, then access will be switched immediately to the mirror copy and if any operation has done on the primary tables those modifications are saved in a journal.
Then after some time whenever the failure is fixed,the amp is brought back online and the primary tables are updated from the journal .
Suppose if we are not using fallback, an amp failure occurs then access must be stopped to those non-fallback tables. Then we need to call field engineer for failure to be fixed or replaced and some intervention would be needed to recreate non-fallback tables.
To update fallback table if non-fallback table was being read and failure occurs then the restoration also requires some intervention to restart the update process.
The usage of disk space is increased (doubled) if a table is fallback turned on. This is due to every row in the primary table will be duplicated into the fallback table.
If we are not using fallback then the performance of sql updates (while all amps are online) , Fastloads and Multiloads can be improved.
Non-fallback can provide less disk space usage and performance improvements. But in summary fallback provides instant recovery and non-fallback need intervention for recovery.
The permspace for each table is distributed differently. Each row of the Primary copy is distributed by the primary index. Each row of the Fallback copy is distributed by the primary index and to a different amp within the same cluster. This means that the permspace used by each copy is distributed differently.
The distribution efficiency of a table is measured by the following:
Average Permspace across all amps
-------------------------------------------------
Maximum Permspace on 1 amp
The distribution efficiency of the total of the Primary and Fallback permspaces is not a true indicator of the distribution efficiency of the Primary copy. In fact, the distribution efficiency of the Primary copy could be perfect (100%), but the distribution efficiency of the Fallback copy could look bad (70%).
Therefore, the only way to truly measure the distribution efficiency of a table is to temporarily drop the Fallback copy, and measure the Primary permspace distribution.
No comments:
Post a Comment