hum: batch process obs

[bingyuyap]

This PR batches observation for processing to reduce DB calls to bigtable, which is the main bottle neck of the performance for processing. This causes the buffer to max out utilisation and drop observations before we can even process them.

High level:

1. whenever there is an observation, add to a batch.
2. whenever batch size reaches batchLimit (for now its 100) or every 5 seconds, process the batch

High level of batch processing:

1. If message does not exist in cache, grab from db and put it in cache. If not found in db assume we haven't seen it before and create a new one in cache.
2. Grab all observations related and put in cache. Check current observation against in-cache observations for deduplication. If not exist, add to cache
3. After every batch, flush these data to db and clear cache for next batch.

Currently on logs, without batching, a buffer size of 20,000 is at 99-100%. Which causes buffer to drop observations. With this change, the utilisation of the buffer is at 0-2% with a size of 1024.

[bruce-riley]

Not quite done yet, but here's what I've got so far.