Today, Ning, Huijun, Sanjeev and Maosong discussed about the
bp(back-pressure) today, and here are the notes:

*Two major issues of current bp algorithm*:
1. one single point of bp will stop the whole topology
2. when a topology in bp state, the overall throughput for the whole
topology can reduce

*Some proposals*:
For 1.
- load shedding -- for example, drop tuples in stmgr.
- better bp algo. -- no need to stop the whole topology; can just handle
the slowness to particular instances, for example, spill the buffer to the
disk

For 2.
- Rate control on the source(spout) side. So even in bp state, the overall
throughput can be higher than normal
- Run-time scale-up

Let me know if I missing anything.

-- 
With my best Regards
------------------
Fu Maosong
Twitter Inc.
Mobile: +001-415-244-7520

Reply via email to