Hi Roshan,
Thank you for the detailed explanation.
In our experiments, we are using cgroups. Each of the components will be
allocated certain amount of cpu by using cgroups (and it will not be shared
with other components). In this scenario, we want to conduct some experiments
with and without backpressure (aiming to maximize the allocated cpu
utilization, and allowing the components to execute in speeds unrelated to
others as you mentioned).
Is there a way to disable backpressure by modifying the code? From the code
it seems like the backpressure status is set whenever any tuples are entered
into backpressure queue. The status seems to be communicated to the parent
through netty. Is there a way to modify the code, so that the parent just drops
the tuple at the parent if the backpressure status is set?
thanks for your time and help.
warm regards,
ravi
________________________________
From: Roshan Naik <[email protected]>
Sent: Friday, April 13, 2018 1:03 PM
To: [email protected]
Subject: Re: Disable backpressure in Storm 2.0
For at-most-once semantics you need to disable ACKing by setting acker count to
0. BP is not intended for that.
IMO you need BP for what you are trying to do. Dropping messages under BP will
lead to a situation where each component runs at a speed that is unrelated to
the rest of the topology and this will skew your CPU consumption measurements.
Some of the CPU will get wasted in processing messages that get discarded.
Also this leads to a very degenerate case of at-most once where there is heavy
data loss.
On Thursday, April 12, 2018, 7:40:30 AM PDT, ravi kiran puttaswamy
<[email protected]> wrote:
Thanks Roshan for the detailed answer.
This is for a research project, which assumes a topology configured in an at
most once semantics. Loss of tuples is tolerable. We want to study the
relationship between maximum throughput and the amount of processing power
(mainly CPU) allocated to the topology. We want to ensure that the topology is
using as much of the CPU provided to it. In case of backpressure, the input it
getting throttled. (I am additionally using cgroups to control the amount of
CPU allocated to each component. So storm 2.0 is my only option).
Is it possible to configure storm to simulate at most once semantics, where the
topology is not throttled by back pressure?
Thanks again for your time,
Warm regards,
ravi
________________________________
From: Roshan Naik <[email protected]>
Sent: Thursday, April 12, 2018 1:09 PM
To: [email protected]
Subject: Re: Disable backpressure in Storm 2.0
Short answer:
It is not possible to disable backpressure in 2.0 and you don't want to do it
either.
Long Answer:
That setting applies to the 1.0 backpressure subsystem (zookeeper based) which
is layered on top of the messaging system. Storm 2.0 has a new messaging
subsystem and a very lightweight backpressure model that is tightly integrated
into the messaging subsystem (with no ZK or other external dependencies).
WRT why you don't want to disable it... If you disable backpressure & allow
upstream components to blindly pump out messages, then you are left with two
options to handle a BP situation. To hold the excess messages, you could allow
the internal queues to keep growing in an unbounded fashion, in which case the
worker process will die with an OOM exception relatively quickly. The other
option is to keep the queues bounded and drop messages once queues are full.
Both options are bad options.
In Storm 1.x you could disable the BP system (default) and fallback on
topology.max.spout.pending as an alternative BP model (if acking is enabled).
If ACKing was disabled and BP is also disabled it is easy to crash the workers.
Any reason you are looking to disable BP ?
-roshan
On Wednesday, April 11, 2018, 7:42:29 PM PDT, ravi kiran puttaswamy
<[email protected]> wrote:
Hello all,
I am configuring a storm topology to execute in at-most once semantics
without backpressure.
Can you let me know how to disable backpressure in storm 2.0? The
documentation says the earlier config "topology.backpressue.enable" will be
deprecated and removed from storm 2.0.
thanks and regards,
ravi
