Reduce heap memory usage about sidewinder

[ambud]

After doing a thorough analysis on the heap and objects, this issue boils down to 2 factors:

1. Size of the time series bucket used
2. Buffer (measurement.increment.size)

Size of time series bucket determines how many buckets will be created and the series retention policy describes how many buckets will be preserved in memory after each Sidewinder GC cycle. Tuning the bucket size is important since data can only be GC'ed in bucket chunks i.e. partial bucket GCs can't be performed.

measurement.increment.size
This setting determines how big of a buffer is allocated to a writer by the measurement class on request. The bigger this buffer the more data can be stored, if this buffer overflows but newer data still falls in the same bucket then new buffer chunks will be allocated and a pointer in the bucket list is created.

Therefore the above two settings must be adjusted in tandem based on the frequency and types of data received since that drives the effective compression ratio that can be achieved. Essentially, the bucket should be large enough to not create too many small buffer chunks such that heap runs out while making sure it's not too large that the retention period becomes unreasonable for your use case and the read performance is effected (all data of at least 1 bucket requires a full scan on reads). The increment size should be adjusted such that too large buffers aren't allocated such that compression is resulting in a mostly empty buffer for a bucket of time.