I have an app, that is reading events from Amazon SQS. The problem, I have is that when I deploy a newer application version, it connects to same queue, so there are two stacks - old and new one consuming messages.
I would like to keep old stack consuming say 95% of messages and new only 5% so I can do a live test. When I am confident, new version is fine, I shut down old stack and make new one consume 100% of events.
The only solution, I see right now is to implement some feature on application side, for example some REST endpoint, to control how many SQS messages it should try to read.
However, may be there are some other solutions/tools for this problem. (in fact, there are several applications, so if I can solve this issue without touching all of them, it would be great)
In general how do you deal with new version deployments and reading from SQS?
thanks
Let's say there are two application stacks: S1, which needs to process 90% of the messages, and S2 10%.
This is what they can do:
They'll have two configurations: n_messages_to_get and n_messages_to_process. For S1, the values will be 10 and 9 respectively. For S2, 10 and 1 respectively.
Each will fetch n_messages_to_get from SQS, but only process n_messages_to_process out of them.
You can also think of having this configuration in a database like DynamoDB, so that you don't have to deploy your code in case you need to dial up or down.
Assumptions made:
Both S1 and S2 take approximately same time to process a message.
You can tolerate some deviation in the number of messages processed by both. For e.g., you'll be OK if S1 processes 87% of the messages and S2 13%.
Related
I'm curious if anyone can point me towards greater visibility into how various Beam Runners manage autoscaling. We seem to be experiencing hiccups during both the 'spin up' and 'spin down' phases, and we're left wondering what to do about it. Here's the background of our particular flow:
1- Binary files arrive on gs://, and object notification duly notifies a PubSub topic.
2- Each file requires about 1Min of parsing on a standard VM to emit about 30K records to downstream areas of the Beam DAG.
3- 'Downstream' components include things like inserts to BigQuery, storage in GS:, and various sundry other tasks.
4- The files in step 1 arrive intermittently, usually in batches of 200-300 every hour, making this - we think - an ideal use case for autoscaling.
What we're seeing, however, has us a little perplexed:
1- It looks like when 'workers=1', Beam bites off a little more than it can chew, eventually causing some out-of-RAM errors, presumably as the first worker tries to process a few of the PubSub messages which, again, take about 60 seconds/message to complete because the 'message' in this case is that a binary file needs to be deserialized in gs.
2- At some point, the runner (in this case, Dataflow with jobId 2017-11-12_20_59_12-8830128066306583836), gets the message additional workers are needed and real work can now get done. During this phase, errors decrease and throughput rises. Not only are there more deserializers for step1, but the step3/downstream tasks are evenly spread out.
3-Alas, the previous step gets cut short when Dataflow senses (I'm guessing) that enough of the PubSub messages are 'in flight' to begin cooling down a little. That seems to come a little too soon, and workers are getting pulled as they chew through the PubSub messages themselves - even before the messages are 'ACK'd'.
We're still thrilled with Beam, but I'm guessing the less-than-optimal spin-up/spin-down phases are resulting in 50% more VM usage than what is needed. What do the runners look for beside PubSub consumption? Do they look at RAM/CPU/etc??? Is there anything a developer can do, beside ACK a PubSub message to provide feedback to the runner that more/less resources are required?
Incidentally, in case anyone doubted Google's commitment to open-source, I spoke about this very topic with an employee there yesterday, and she expressed interest in hearing about my use case, especially if it ran on a non-Dataflow runner! We hadn't yet tried our Beam work on Spark (or elsewhere), but would obviously be interested in hearing if one runner has superior abilities to accept feedback from the workers for THROUGHPUT_BASED work.
Thanks in advance,
Peter
CTO,
ATS, Inc.
Generally streaming autoscaling in Dataflow works like this :
Upscale: If the pipeline's backlog is more than a few seconds based on current throughput, pipeline is upscaled. Here CPU utilization does not directly affect the amount of upsize. Using CPU (say it is at 90%), does not help in answering the question 'how many more workers are required'. CPU does affect indirectly since pipelines fall behind when they they don't enough CPU thus increasing backlog.
Downcale: When backlog is low (i.e. < 10 seconds), pipeline is downcaled based on current CPU consumer. Here, CPU does directly influence down size.
I hope the above basic description helps.
Due to inherent delays involved in starting up new GCE VMs, the pipeline pauses for a minute or two during resizing events. This is expected to improve in near future.
I will ask specific questions about the job you mentioned in description.
I am looking to run a service that will be consuming messages that are placed into an SQS queue. What is the best way to structure the consumer application?
One thought would be to create a bunch of threads or processes that run this:
def run(q, delete_on_error=False):
while True:
try:
m = q.read(VISIBILITY_TIMEOUT, wait_time_seconds=MAX_WAIT_TIME_SECONDS)
if m is not None:
try:
process(m.id, m.get_body())
except TransientError:
continue
except Exception as ex:
log_exception(ex)
if not delete_on_error:
continue
q.delete_message(m)
except StopIteration:
break
except socket.gaierror:
continue
Am I missing anything else important? What other exceptions do I have to guard against in the queue read and delete calls? How do others run these consumers?
I did find this project, but it seems stalled and has some issues.
I am leaning toward separate processes rather than threads to avoid the the GIL. Is there some container process that can be used to launch and monitor these separate running processes?
There are a few things:
The SQS API allows you to receive more than one message with a single API call (up to 10 messages, or up to 256k worth of messages, whichever limit is hit first). Taking advantage of this feature allows you to reduce costs, since you are charged per API call. It looks like you're using the boto library - have a look at get_messages.
In your code right now, if processing a message fails due to a transient error, the message won't be able to be processed again until the visibility timeout expires. You might want to consider returning the message to the queue straight away. You can do this by calling change_visibility with 0 on that message. The message will then be available for processing straight away. (It might seem that if you do this then the visibility timeout will be permanently changed on that message - this is actually not the case. The AWS docs state that "the visibility timeout for the message the next time it is received reverts to the original timeout value". See the docs for more information.)
If you're after an example of a robust SQS message consumer, you might want to check out NServiceBus.AmazonSQS (of which I am the author). (C# - sorry, I couldn't find any python examples.)
Somebody told me that simple_one_for_one is very useful for chat applications, because each chat client is a server process (gen_server). Is this right?
And I wonder why do we need it? Why not create only one center server (gen_server) to handle all chat client communication? Because maybe the number of chat clients is very large so only one server couldn't handle fast, make the system slow down?
I think maybe creating too many servers like simple_one_for_one may take too much system resource. I'm a new OTP guy, so I really need explanation about this point.
Yes, the idea is that you would have a process (gen_server) per client.
This lets you isolate failure of one client from another.
If you had everyone in a single process, you have to be very careful to handle all the things that might go wrong and crash you process (thus, disconnecting all your clients).
With one process per client, you can code for the happy path and just let it crash when things go wrong. Worst case is you drop a single client.
Processes are fairly cheap (nothing like creating threads). On a modern machine you can have millions.
If your user base is in the many millions, I'm sure you'd end up with more than one server anyway. So something that can easily scale to the hundreds of thousands to low millions on a box is plenty.
I'm introducing 24/7 monitoring for our systems. To avoid unnecessary pages in the middle of the night I want Nagios to NOT page me, if only one or two of the service checks fail, as this won't have any impact on users: The other servers run the same service and the impact on users is almost zero, so fixing the problem has time until the next day.
But: I want to get paged if too many of the checks fail.
For example: 50 servers run the same service, 2 fail -> I can still sleep.
The service fails on 15 servers -> I get paged because the impact is getting to high.
What I could do is add a lot (!) of notification dependencies that only trigger if numerous hosts are down. The problem: Even though I can specify to get paged if 15 hosts are down, I still have to define exactly which hosts need to be down for this alert to be sent. I rather want to specify that if ANY 15 hosts are down a page is made.
I'd be glad if somebody could help me with that.
Personally I'm using Shinken which has business rules just for that. Shinken is backward compatible with Nagios, so it's easy to drop your nagios configuration into shinken.
It seems there is a similar addon for nagios Nagios Business Process Intelligence Addon, but I'm not having experience with this addon.
We have an app that uses Cloudant as a remote server. Nevertheless, Cloudant is not completely compatible with TouchDB's continuous replications from previous experience. So our alternative for now is to trigger manually one-shot replications at a fixed frequency. Nevertheless, we would like to know if that approach is going to cost us more money than continuous replications, since continuous replications use longpoll and doesn't need to query the server often. In other words, does one-shot pull replications with Cloudant as the target cost us a GET request?
Thank you,
Paul
I think the issue you refer to is [1].
Cloudant's replication is 100% compatible with CouchDB. In this
instance, TouchDB's logs indicate the iOS network stack passed
on incomplete JSON to TouchDB. It's not clear who was to blame
in this case for the replication failure.
[1] https://github.com/couchbaselabs/TouchDB-iOS/issues/241
For the cost question, a one-shot pull replication will result in a GET to the _changes
feed each time it happens, plus the other requests required to
replicate. This _changes request will be counted as a light
HTTP request against your Cloudant account.
However, whether this works out as more or fewer requests overall
depends on the number of changes coming down from the remote server.
It's also important to remember that the number of _changes calls are very small
relative to the number of other calls involved (e.g., getting the
content of the changes themselves and particularly if there are many
attachments).
While this question is specific to TouchDB, and I mention specific
behaviours of that codebase, this answer deals with the requests involved
in replication between any two systems speaking the CouchDB replication
protocol[2].
[2] http://www.dataprotocols.org/en/latest/couchdb_replication.html
Let's take a contrived example: 1 update per 10 second window to
the source database for the replication, where a TouchDB database
is the target. Let's take a 5 minute poll vs. a continuous replication.
For simplicity of call-counting, let's also take attachments out of the
picture. We'll also assume the device has a constant network connection.
For the continuous case, every 10s TouchDB will receive an update in
the _changes feed. This causes the longpoll connection to close.
TouchDB then runs through the changes, requesting the updates from the
source database; one or more GET requests on the remote server. While
this is happening, TouchDB has to open up another longpoll request
to _changes. So in a five minute period, you'd end up with perhaps
30 calls to _changes, plus all the calls to get documents and record
checkpoints.
Compare this with a one-shot replication every five minutes. You'd
receive notification of the 30 updates in one _changes feed call.
TouchDB implements an optimisation[3] whereby it will call _all_docs
to get updated documents for 1- revs, so you might end up with a single
call to get all 30 documents (not possible in the continuous case as
you've received a single change). Then you've the checkpoint documents
to record. At best fewer than 5 HTTP calls, at most about a third of
the continuous case as you've avoided extra _changes requests.
[3] https://github.com/couchbaselabs/TouchDB-iOS/wiki/Replication-Algorithm#performance
It comes down to the frequency of updates you expect to the source
database. One-shot replication is likely to provide a smoother price
curve as you're in better control of the number of requests you make.
A further question is how often connections will drop because of the
network disconnects which happen regularly with mobile devices.
TouchDB's continuous replications will fire back up each time the
user comes on line (if added via the _replicator database). This is a
further source of unpredictable costs.
However, the benefits from more immediate visibility of changes may
certainly be worth the uncertainty.