Slowly increasing memory usage of Dask Sheduler - dask

Question

I'm running a test:
client = Client('127.0.0.1:8786')
def x(i):
return {}
while True:
start = time.time()
a = client.submit(randint(0,1000000))
res = a.result()
del a
end = time.time()
print("Ran on %s with res %s" % (end-start, res))
client.shutdown()
del client
I used it (with more code) to get an estimate of my queries performance. But for this example I've removed all things I could think of.
The above code leaks roughly 0.1 MB per second, which I would guesstimate to roughly 0.3MB per 1000 calls.
Am I doing something wrong in my code?

Answer 1

My python debugging skills are a bit rusty (and with a bit I mean I last used objgraph on Orbited (the precursor to websockets) in 2009 https://pypi.python.org/pypi/orbited) but from what I can see, checking to number of references before and after:
Counting objects in the scheduler, before and after using objgraph.show_most_common_types()
| What | Before | After | Diff |
|-------------+------------------+--------|---------+
| function | 33318 | 33399 | 81 |
| dict | 17988 | 18277 | 289 |
| tuple | 16439 | 28062 | 11623 |
| list | 10926 | 11257 | 331 |
| OrderedDict | N/A | 7168 | 7168|
It's not a huge number of RAM in any case, but digging deeper I found that t scheduler._transition_counter is 11453 and scheduler.transition_log is filled with:
('x-25ca747a80f8057c081bf1bca6ddd481', 'released', 'waiting',
OrderedDict([('x-25ca747a80f8057c081bf1bca6ddd481', 'processing')]), 4121),
('x-25ca747a80f8057c081bf1bca6ddd481', 'waiting', 'processing', {}, 4122),
('x-25cb592650bd793a4123f2df39a54e29', 'memory', 'released', OrderedDict(), 4123),
('x-25cb592650bd793a4123f2df39a54e29', 'released', 'forgotten', {}, 4124),
('x-25ca747a80f8057c081bf1bca6ddd481', 'processing', 'memory', OrderedDict(), 4125),
('x-b6621de1a823857d2f206fbe8afbeb46', 'released', 'waiting', OrderedDict([('x-b6621de1a823857d2f206fbe8afbeb46', 'processing')]), 4126)
First error on my part
Which of course led me to realise the first error on my part was not configuring transition-log-length.
After setting configuration transition-log-length to 10:
| What | Before | After | Diff |
| ---------------+----------+--------+---------|
| function | 33323 | 33336 | 13 |
| dict | 17987 | 18120 | 133 |
| tuple | 16530 | 16342 | -188 |
| list | 10928 | 11136 | 208 |
| _lru_list_elem | N/A | 5609 | 5609 |
A quick google found that _lru_list_elem is made by #functools.lru_cache which in turn is in invoked in key_split (in distributed/utils.py)
Which is the LRU cache, of up to 100 000 items.
Second try
Based on the code it appears as Dask should climb up to roughly 10k _lru_list_elem
After running my script again and watching the memory it climbs quite fast up until I approach 100k _lru_list_elem, afterwards it stops climbing almost entirely.
This appears to be the case, since it pretty much flat-lines after 100k
So no leak, but fun to get hands dirty on Dask source code and Python memory profilers

Answer 2

For diagnostic, logging, and performance reasons the Dask scheduler keeps records on many of its interactions with workers and clients in fixed-sized deques. These records do accumulate, but only to a finite extent.
We also try to ensure that we don't keep around anything that would be too large.
Seeing memory use climb up until a nice round number like what you've seen and then stay steady seems to be consistent with this.