Details
-
New Feature
-
Resolution: Fixed
-
Major
-
None
-
None
-
Security Level: Public
-
None
Description
When doing things like bulk loading, there are times we have multiple workload profiles. Interactive applications using the same client and same server need "room to move" and should always be fast, while other workloads should throttle themselves.
The intent here would be to add a feature such that:
for every N operations, pull stats from the nodes
if the amount of memory used is below high_wat, don't do anything
else if the amount of memory used is 10% of the remaining memory above high_wat before max memory usage, insert a sleep of Y milliseconds before each operation, and log the backoff at a debug level, check stats again ever M operations
else if the amount of memory used is greater than 10% above that level above high_wat, insert a sleep of Z milliseconds before each operation and log the backoff at an info level, check stats again every O operations
Turning this on or off should be done via a properties file or just simple -D args to the JVM, either one is okay.
N, M, O and Y, Z should all be tuneable, either via a properties file or a define. All of them should be optional.
Default values should be:
N = 10000
M = N/100
O = M/10
Y = 1ms
Z = Y * 3