I am using an F4 instance (because of memory needs) with automatic scheduling to do some background processing. It is run from a task queue. It takes 40s to 60s to complete each invocation. Because of the high memory needs, each instance should only handle one request at a time.
The action that needs to be done is not urgent. If it doesn't get scheduled for 30 minutes that isn't a problem. Even 60 minutes is acceptable and I'd rather make use of that time rather than spin up more instances. However, if the service gets popular and the is getting more than 60 requests an hour I want to spin up more instances to make sure there isn't more than a 60 minute wait.
I am having trouble figuring out how to configure the instance and queue parameters to keep my costs down but be able to scale in that way. My initial thought was something like this:
<queue>
<name>non-urgent-queue</name>
<target>slow-service</target>
<rate>1/m</rate>
<bucket-size>1</bucket-size>
<max-concurrent-requests>1</max-concurrent-requests>
</queue>
<automatic-scaling>
<min-idle-instances>0</min-idle-instances>
<max-idle-instances>0</max-idle-instances>
<min-pending-latency>20m</min-pending-latency>
<max-pending-latency>1h</max-pending-latency>
<max-concurrent-requests>1</max-concurrent-requests>
</automatic-scaling>
First of all those latency settings are invalid, but I can't find documentation on the valid range or units. Can anyone direct me to that info?
Secondly, if I understand the queue settings correctly, this configuration would limit it to 60 invocations an hour getting to the service, even if the task queue had 60+ jobs waiting.
Thanks for your help!
Indeed, throttling at the queue level basically defeats the ability to scale when needed. So you can't use the <rate> in the queue configuration at the values you have right now, you need to use the value matching the maximum rate you're willing to accept (with you max number of instances running simultaneously):
the max rate of requests that can go through the queue being limited at 1/min means you can't scale above 60/h
the <bucket-size> set at 1 means no peaks above the rate can be handled (as soon as one task starts the token bucket empties).
the <max-concurrent-requests> set at 1 will basically prevent multiple instances dealing simultaneouly with the queued workload. They may be started by the autoscaler because of the request latencies, but they won't be able to help since only one queue task can be handled at a time.
In the <automatic-scaling> section the <max-concurrent-requests> set to 1 is good - this ensures no instance handles more than 1 request at a time - which is what you want.
The bad news is that the max values for the latencies appear to be 15s. At least when using the app.yaml config for python (but I think it's unlikely for that to differ across language sandboxes):
Error 400: --- begin server output ---
automatic_scaling.min_pending_latency (30s), must be in the range [0.010000s,15.000000s].
--- end server output ---
and
Error 400: --- begin server output ---
automatic_scaling.max_pending_latency (60s), must be in the range [0.010000s,15.000000s].
--- end server output ---
Which probably also explains why your 5m and 1h values aren't accepted - I used 30s and 60s and got the above errors.
This means you won't be able to use the autoscaling parameters to tune such a slow-moving processing like you desire.
The only alternative I can think of is to have 2 queues:
a fast one feeding just trigger tasks for the slow-service jobs, but which your service intercepts and saves in the datastore. Maybe performed by some faster service (you don't want these stuck behind a slow-service job execution as it can cause unnecessary instance launching. Maybe, depending on the rest of your implementation, you can replace this queue completely with just storing the job info in the datastore instead of enqueing tasks in the fast queue.
a slow one for the actual slow-service job execution tasks
You'd also have a cron job executing once a minute, checking how many triggers are pending in the datastore, decide how much to scale and enqueue the corresponding number of slow-service job tasks in the slow queue. The autoscaler would simply bring up the corresponding number of instances (if needed). Low latency autoscaling configs would be desirable in this case - you already decided how you want your app to scale.
This is how I ended up doing it. I use a slow queue and a fast queue configured like this:
<queue>
<name>slow-queue</name>
<target>pdf-service</target>
<rate>2/m</rate>
<bucket-size>1</bucket-size>
<max-concurrent-requests>1</max-concurrent-requests>
</queue>
<queue>
<name>fast-queue</name>
<target>pdf-service</target>
<rate>10/m</rate>
<bucket-size>1</bucket-size>
<max-concurrent-requests>5</max-concurrent-requests>
</queue>
The max-concurrent-requests in the slow queue ensures only one task will run at a time, so there will only be one instance active.
Before I post to the slow queue I check to see how many items are already on the queue. The result may not be totally reliable, but for my purposes it is sufficient. In java:
QueueStatistics queueStats = queue.fetchStatistics();
if(queueStats.getNumTasks()<30) {
//post to slow queue
} else {
//post to fast queue
}
So when my slow queue gets too full, I post to the fast queue which allows concurrent requests.
The instance is configured like this:
<automatic-scaling>
<min-idle-instances>0</min-idle-instances>
<max-idle-instances>automatic</max-idle-instances>
<min-pending-latency>15s</min-pending-latency>
<max-pending-latency>15s</max-pending-latency>
<max-concurrent-requests>1</max-concurrent-requests>
</automatic-scaling>
So it will create new instances as slowly as possible (15s is the max latency) and make sure only one process runs on an instance at a time.
With this configuration I'll have a max of 6 instances at a time but that should do about 500/hr. I could increase the rate and concurrent requests to do more.
The negative of this solution is an element of unfairness. Under heavy load, some tasks will be stuck in the slow queue while others will get processed more quickly in the fast queue.
Because of that, I have decreased the max items on the slow queue to 13 so the unfairness won't be so extreme, maybe a 10 minute wait for jobs that go to the slow queue when it is full.
Related
I have a push task queue and each of my jobs consists of multiple similar TaskQueue tasks. Each of these tasks takes less than a second to finish and can add new tasks to the queue (they should be also completed to consider the job finished). Task results are written to a DataStore.
The goal is to understand when a job has finished, i.e. all of its tasks are completed.
Writes are really frequent and I can't store the results inside one entity group. Is there a good workaround for this?
In a similar context I used a scheme based on memcache, which doesn't have a significant write rate limitation as datastore entity groups:
each job gets a unique memcache key associated with it, which it passes to each of subsequent execution tasks it may enqueue
every execution task updates the memcache value corresponding to the job key with the current timestamp and also enqueues a completion check task, delayed with an idle timeout value, large enough to declare the job completed if elapsed.
every completion check task compares the memcache value corresponding to the job key against the current timestamp:
if the delta is less than the idle timeout it means the job is not complete (some other task was executed since this completion check task was enqueued, hence some other completion check task is in the queue)
otherwise the job is completed
Note: the idle timeout should be larger than the maximum time a task might spend in the queue.
In my app, I need to send emails to a large number of users when an event happens. I'd like to send those emails out gradually instead of all at once. For clarity in explaining, let's say I need to send out emails to 10,000 users.
I currently do this with a task queue with a maximum rate of 1 task/second. I enqueue 10,000 tasks in batches, and the emails get sent out at a rate of 1/second.
I'd like to change this to using an ETA for the tasks instead of limiting the task queue to a maximum rate. Conceptually it would be like this (except that task submission would be batched):
now = datetime.utcnow()
for i, email in enumerate(email_list):
eta = now + datetime.timedelta(seconds=i)
deferred.defer(send_email, email, _eta=eta)
Before implementing a change like this, I'd like to have some confidence that GAE can do this efficiently.
If I have 10,000 tasks in a task queue, each with a different ETA, will the GAE task queue be able to efficiently monitor all the tasks and start them at approximately (the precise ETA isn't important) the appropriate times? I don't know what algorithm Google uses for this.
EDIT:
Imagine if you inserted a billion tasks in a single day each with an ETA. How would GAE monitor those tasks to make sure they got fired off at the right time? Polling all the tasks at some time interval (e.g., every minute) would be a terrible solution. Perhaps GAE uses some kind of priority queue. It would be nice to have some confidence that GAE has implemented an algorithm that will scale for a lot of tasks with an ETA.
With the stated daily quota of 10 billion tasks one would think they should be able to handle 10,000 of them :)
In my current project I'm also sending ~10,000 emails (SendGrid) with tasks & _eta (although in batches of 25) which works fine so far...
In the current infrastructure, the logic can be a little batchy when the throughput is significantly below the configured rate. Queues prepare tasks 5s in advance but processing can slow down if there are no tasks in a given 5s window.
It should work in general, but you might see a pattern of delays of up to 20s followed by bursts.
At a total throughput of 1B tasks/day, you would probably want to split to run over 40 queues at a rate of around 300 tasks/sec/queue. With a rate that steady, delays would be uncommon.
I'm trying to solve the following problem:
I have a series of "tasks" which I would like to execute
I have a fixed number of workers to execute these workers (since they call an external API using urlfetch and the number of parallel calls to this API is limited)
I would like for these "tasks" to be executed "as soon as possible" (ie. minimum latency)
These tasks are parts of larger tasks and can be categorized based on the size of the original task (ie. a small original task might generate 1 to 100 tasks, a medium one 100 to 1000 and a large one over 1000).
The tricky part: I would like to do all this efficiently (ie. minimum latency and use as many parallel API calls as possible - without getting over the limit), but at the same time try to prevent a large number of tasks generated from "large" original tasks to delay the tasks generated from "small" original tasks.
To put it an other way: I would like to have a "priority" assigned to each task with "small" tasks having a higher priority and thus prevent starvation from "large" tasks.
Some searching around doesn't seem to indicate that anything pre-made is available, so I came up with the following:
create three push queues: tasks-small, tasks-medium, tasks-large
set a maximum number of concurrent request for each such that the total is the maximum number of concurrent API calls (for example if the max. no. concurrent API calls is 200, I could set up tasks-small to have a max_concurrent_requests of 30, tasks-medium 60 and tasks-large 100)
when enqueueing a task, check the no. pending task in each queue (using something like the QueueStatistics class), and, if an other queue is not 100% utilized, enqueue the task there, otherwise just enqueue the task on the queue with the corresponding size.
For example, if we have task T1 which is part of a small task, first check if tasks-small has free "slots" and enqueue it there. Otherwise check tasks-medium and tasks-large. If none of them have free slots, enqueue it on tasks-small anyway and it will be processed after the tasks added before it are processed (note: this is not optimal because if "slots" free up on the other queues, they still won't process pending tasks from the tasks-small queue)
An other option would be to use PULL queue and have a central "coordinator" pull from that queue based on priorities and dispatch them, however that seems to add a little more latency.
However this seems a little bit hackish and I'm wondering if there are better alternatives out there.
EDIT: after some thoughts and feedback I'm thinking of using PULL queue after all in the following way:
have two PULL queues (medium-tasks and large-tasks)
have a dispatcher (PUSH) queue with a concurrency of 1 (so that only one dispatch task runs at any time). Dispatch tasks are created in multiple ways:
by a once-a-minute cron job
after adding a medium/large task to the push queues
after a worker task finishes
have a worker (PUSH) queue with a concurrency equal to the number of workers
And the workflow:
small tasks are added directly to the worker queue
the dispatcher task, whenever it is triggered, does the following:
estimates the number of free workers (by looking at the number of running tasks in the worker queue)
for any "free" slots it takes a task from the medium/large tasks PULL queue and enqueues it on a worker (or more precisely: adds it to the worker PUSH queue which will result in it being executed - eventually - on a worker).
I'll report back once this is implemented and at least moderately tested.
The small/medium/large original task queues won't help much by themselves - once the original tasks are enqueued they'll keep spawning worker tasks, potentially even breaking the worker task queue size limit. So you need to pace/control enqueing of the original tasks.
I'd keep track of the "todo" original tasks in the datastore/GCS and enqueue these original tasks only when the respective queue size is sufficiently low (1 or maybe 2 pending jobs), from either a recurring task, a cron job or a deferred task (depending on the rate at which you need to perform the original task enqueueing) which would implement the desired pacing and priority logic just like a push queue dispatcher, but without the extra latency you mentioned.
I have not used pull queues, but from my understanding they could suit your use-case very well. Your could define 3 pull queues, and have X workers all pulling tasks from them, first trying the "small" queue then moving on to "medium" if it is empty (where X is your maximum concurrency). You should not need a central dispatcher.
However, then you would be left to pay for X workers even when there are no tasks (or X / threadsPerMachine?), or scale them down & up yourself.
So, here is another thought: make a single push queue with the correct maximum concurrency. When you receive a new task, push its info to the datastore and queue up a generic job. That generic job will then consult the datastore looking for tasks in priority order, executing the first one it finds. This way a short task will still be executed by the next job, even if that job was already enqueued from a large task.
EDIT: I now migrated to a simpler solution, similar to what #eric-simonton described:
I have multiple PULL queues, one for each priority
Many workers pull on an endpoint (handler)
The handler generates a random number and does a simple "if less than 0.6, try first the small queue and then the large queue, else vice-versa (large then small)"
If the workers get no tasks or an error, they do semi-random exponential backoff up to maximum timeout (ie. they start pulling every 1 second and approximately double the timeout after each empty pull up to 30 seconds)
This final point is needed - amongst other reasons - because the number of pulls / second from a PULL queue is limited to 10k/s: https://cloud.google.com/appengine/docs/python/taskqueue/overview-pull#Python_Leasing_tasks
I implemented the solution described in the UPDATE:
two PULL queues (medium-tasks and large-tasks)
a dispatcher (PUSH) queue with a concurrency of 1
a worker (PUSH) queue with a concurrency equal to the number of workers
See the question for more details. Some notes:
there is some delay in task visibility due to eventual consistency (ie. the dispatchers tasks sometimes don't see the tasks from the pull queue even if they are inserted together) - I worked around by adding a countdown of 5 seconds to the dispatcher tasks and also adding a cron job that adds a dispatcher task every minute (so if the original dispatcher task doesn't "see" the task from the pull queue, an other will come along later)
made sure to name every task to eliminate the possibility of double-dispatching them
you can't lease 0 items from the PULL queues :-)
batch operations have an upper limit, so you have to do your own batching over the batch taskqueue calls
there doesn't seem to be a way to programatically get the "maximum parallelism" value for a queue, so I had to hard-code that in the dispatcher (to calculate how many more tasks it can schedule)
don't add dispatcher tasks if they are already some (at least 10) in the queue
We have an application that takes some input from a user and makes ~50 RPC calls. Each call takes around 4-5 minutes.
In the backend we are using a push queue and enqueuing each of these 50 calls as tasks. This is our queue spec:
queue:
- name: some-name
rate: 500/s
bucket_size: 100
max_concurrent_requests: 500
My understanding is that all 50 requests should be run in parallel, and thus all of them should be complete in 4-5 minutes. But what's actually happening is that only around ~15 of these requests are returning results, while the rest cross the 10 min limit and time out. Another thing to note is that this seems to work fine if we bring down the number of requests to < 10.
There's always the possibility that the requests that timed out did so because the RPC response actually took that long. But what I wanted to confirm is :
My understanding of the tasks running in parallel is correct.
Our queue config and the number of tasks we're enqueuing has nothing to do with these requests timing out.
Are these correct ?
(1) Parallel execution
Yes, tasks can be executed in parallel (up to 500 in your case), but in push queues, your app has no control in which particular order the tasks in a push queue are executed and no direct control how many tasks are executed at once. (Your app can control in which sequence tasks are added to a queue though, see the pattern in (2) below)
App Engine uses certain factors to decide how fast and which tasks are executed, especially the queue configuration and also the scaling configuration (e.g. in app.yaml). Since you pay for every first 15 minutes of an instance, it could get very expensive to really have 50 instances launched, then idling for 15 minutes before shutting them down (until the next request). In this regard, the mechanism that spawns new instances is a little smarter, whether it is HTTP requests by users or task queues.
(2) Request time outs
Yes, it is very unlikely that the enqueuing has anything to do with these request time outs. Unless the time-outs are an unintentional side-effect of the wrong assumption that a particular task was executed before.
In order to avoid request time outs in general, it makes sense to split a task into multiple tasks. For example, if you have a task do_foo and those executions exceed the time outs frequently (or memory limits), you could instead have do_foo load off work to other tasks that will do the actual jobs.
For some migration tasks I use this pattern in a linear / sequential way. E.g. classmethod do_foo just queries entities of a certain kind (ordered by creation timestamp for example), maybe filtered, by page (e.g. 50 in transactions with ancestor). It does some writes to the entities first, and only at the very end after successful commit it creates a new transactional do_foo task with cursor parameter to the next page, eventually with a countdown of 1 sec to avoid transaction errors. The next execution of do_foo will continue with the next page (of course only after the task with the previous page completed).
Depending on the nature of the tasks, you could alternatively have each task fan out into multiple tasks per execution, e.g. do_foo triggers do_bar, do_something and do_more. Also note that up to five tasks can be created transactionally inside a transaction.
I have a bit of a strange problem. I have a module running on gae that puts a whole lot of little tasks on the default task queue. The tasks access the same ndb module. Each task accesses a bunch of data from a few different tables then calls put.
The first few tasks work fine but as time continues I start getting these on the final put:
suspended generator _put_tasklet(context.py:358) raised TransactionFailedError(too much contention on these datastore entities. please try again.)
So I wrapped the put with a try and put in a randomised timeout so it retries a couple of times. This mitigated the problem a little, it just happens later on.
Here is some pseudocode for my task:
def my_task(request):
stuff = get_ndb_instances() #this accessed a few things from different tables
better_stuff = process(ndb_instances) #pretty much just a summation
try_put(better_stuff)
return {'status':'Groovy'}
def try_put(oInstance,iCountdown=10):
if iCountdown<1:
return oInstance.put()
try:
return oInstance.put()
except:
import time
import random
logger.info("sleeping")
time.sleep(random.random()*20)
return oInstance.try_put(iCountdown-1)
Without using try_put the queue gets about 30% of the way through until it stops working. With the try_put it gets further, like 60%.
Could it be that a task is holding onto ndb connections after it has completed somehow? I'm not making explicit use of transactions.
EDIT:
there seems to be some confusion about what I'm asking. The question is: Why does ndb contention get worse as time goes on. I have a whole lot of tasks running simultaneously and they access the ndb in a way that can cause contention. If contention is detected then a randomy timed retry happens and this eliminates contention perfectly well. For a little while. Tasks keep running and completing and the more that successfully return the more contention happens. Even though the processes using the contended upon data should be finished. Is there something going on that's holding onto datastore handles that shouldn't be? What's going on?
EDIT2:
Here is a little bit about the key structures in play:
My ndb models sit in a hierarchy where we have something like this (the direction of the arrows specifies parent child relationships, ie: Type has a bunch of child Instances etc)
Type->Instance->Position
The ids of the Positions are limited to a few different names, there are many thousands of instances and not many types.
I calculate a bunch of Positions and then do a try_put_multi (similar to try_put in an obvious way) and get contention. I'm going to run the code again pretty soon and get a full traceback to include here.
Contention will get worse overtime if you continually exceed the 1 write/transaction per entity group per second. The answer is in how Megastore/Paxo work and how Cloud Datastore handles contention in the backend.
When 2 writes are attempted at the same time on different nodes in Megastore, one transaction will win and the other will fail. Cloud Datastore detects this contention and will retry the failed transaction several times. Usually this results in the transaction succeeding without any errors being raised to the client.
If sustained writes above the recommended limit are being attempted, the chance that a transaction needs to be retried multiple times increases. The number of transactions in an internal retry state also increases. Eventually, transactions will start reaching our internal retry limit and will return a contention error to the client.
Randomized sleep method is an incorrect way to handle error response situations. You should instead look into exponential back-off with jitter (example).
Similarly, the core of your problem is a high write rate into a single entity group. you should look into whether the explicit parenting is required (removing it if not), or if you should shard the entity group in some manner that makes sense according to your queries and consistency requirements.