I have a list of numbers. This list is stored in two ways: either as an in-memory python object, or as a redis list (redis set up in the same server).
I'm comparing the time it takes to retrieve these two lists, using python's timeit. Here's what I do in the python shell:
import timeit
import redis
POOL = redis.ConnectionPool(host='127.0.0.1',port=6379,db=0)
my_server = redis.Redis(connection_pool=POOL)
print min(timeit.Timer('pylist1 = my_server.lrange("nums:5",0,-1)', setup='from __main__ import my_server').repeat(7,1000))
This gives me a time of 1.92341279984.
Next, I time the in-memory python object like so:
pylist = my_server.lrange("nums:5",0,-1)
print min(timeit.Timer('pylist2 = pylist',setup='from __main__ import pylist').repeat(7,1000))
This gives me a time of 4.29153442383e-05. I.e. it seems to be ~45K times faster than retrieving the same list from redis.
My question is this: is my comparison approach correct? I.e., am I accurately simulating retrieval from redis vs retrieval from memory? This is a huge performance boost for the use case I have in mind, but before I implement this, just want to be sure I didn't fudge the benchmarking.
In the comparison you've put up here, you're basically just measuring how long Python takes to bind a new name to a value in the second case. So it doesn't surprise me that this is vastly faster than communicating with a different process (Redis). I guess what surprises me is that you would consider getting a value from Redis if the option exists simply to keep it in memory.
So, you need to be more clear about why you are using Redis for this in the first place. It will always be slower than in-process memory, no benchmark needed for that. You need to ask "why am I not just using Python lists and dictionaries"? There are several valid answers: your data is too large to fit into memory, you require the cache-specific features like allowing values to disappear after a while, or you want to use it for IPC, or persistence. Once you know the answer here, that will inform the benchmarking you want to do. And the question will be more like "How do I obtain the benefits/features I have listed above for the least performance penalty". Redis may not be the only answer. You may consider shelf for persistence, or perhaps even a full-on relational database or Mongo or whatever.
In short, once you have a good idea of why, the how often solves itself.
Related
I'm currently running into an issue in integrating ElasticSearch and MongoDB. Essentially I need to convert a number of Mongo Documents into searchable documents matching my ElasticSearch query. That part is luckily trivial and taken care of. My problem though is that I need this to be fast. Faster than network time, I would really like to be able to index around 100 docs/second, which simply isn't possible with network calls to Mongo.
I was able to speed this up a lot by using ElasticSearch's bulk indexing, but that's only half of the problem. Is there any way to either bundle reads or cache a collection (a manageable part of a collection, as this collection is larger than I would like to keep in memory) to help speed this up? I was unable to really find any documentation about this, so if you can point me towards relevant documentation I consider that a perfectly acceptable answer.
I would prefer a solution that uses Pymongo, but I would be more than happy to use something that directly talks to MongoDB over requests or something similar. Any thoughts on how to alleviate this?
pymongo is thread safe, so you can run multiple queries in parallel. (I assume that you can somehow partition your document space.)
Feed the results to a local Queue if processing the result needs to happen in a single thread.
I doubt this is even possible, but here is the problem and proposed solution (the feasibility of the proposed solution is the object of this question):
I have some "global data" that needs to be available for all requests. I'm persisting this data to Riak and using Redis as a caching layer for access speed (for now...). The data is split into about 30 logical chunks, each about 8 KB.
Each request is required to read 4 of these 8KB chunks, resulting in 32KB of data read in from Redis or Riak. This is in ADDITION to any request-specific data which would also need to be read (which is quite a bit).
Assuming even 3000 requests per second (this isn't a live server so I don't have real numbers, but 3000ps is a reasonable assumption, could be more), this means 96KBps of transfer from Redis or Riak in ADDITION to the already not-insignificant other calls being made from the application logic. Also, Python is parsing the JSON of these 8KB objects 3000 times every second.
All of this - especially Python having to repeatedly deserialize the data - seems like an utter waste, and a perfectly elegant solution would be to just have the deserialized data cached in an in-memory native object in Python, which I can refresh periodically as and when all this "static" data becomes stale. Once in a few minutes (or hours), instead of 3000 times per second.
But I don't know if this is even possible. You'd realistically need an "always running" application for it to cache any data in its memory. And I know this is not the case in the nginx+uwsgi+python combination (versus something like node) - python in-memory data will NOT be persisted across all requests to my knowledge, unless I'm terribly mistaken.
Unfortunately this is a system I have "inherited" and therefore can't make too many changes in terms of the base technology, nor am I knowledgeable enough of how the nginx+uwsgi+python combination works in terms of starting up Python processes and persisting Python in-memory data - which means I COULD be terribly mistaken with my assumption above!
So, direct advice on whether this solution would work + references to material that could help me understand how the nginx+uwsgi+python would work in terms of starting new processes and memory allocation, would help greatly.
P.S:
Have gone through some of the documentation for nginx, uwsgi etc but haven't fully understood the ramifications per my use-case yet. Hope to make some progress on that going forward now
If the in-memory thing COULD work out, I would chuck Redis, since I'm caching ONLY the static data I mentioned above, in it. This makes an in-process persistent in-memory Python cache even more attractive for me, reducing one moving part in the system and at least FOUR network round-trips per request.
What you're suggesting isn't directly feasible. Since new processes can be spun up and down outside of your control, there's no way to keep native Python data in memory.
However, there are a few ways around this.
Often, one level of key-value storage is all you need. And sometimes, having fixed-size buffers for values (which you can use directly as str/bytes/bytearray objects; anything else you need to struct in there or otherwise serialize) is all you need. In that case, uWSGI's built-in caching framework will take care of everything you need.
If you need more precise control, you can look at how the cache is implemented on top of SharedArea and do something customize. However, I wouldn't recommend that. It basically gives you the same kind of API you get with a file, and the only real advantages over just using a file are that the server will manage the file's lifetime; it works in all uWSGI-supported languages, even those that don't allow files; and it makes it easier to migrate your custom cache to a distributed (multi-computer) cache if you later need to. I don't think any of those are relevant to you.
Another way to get flat key-value storage, but without the fixed-size buffers, is with Python's stdlib anydbm. The key-value lookup is as pythonic as it gets: it looks just like a dict, except that it's backed up to an on-disk BDB (or similar) database, cached as appropriate in memory, instead of being stored in an in-memory hash table.
If you need to handle a few other simple types—anything that's blazingly fast to un/pickle, like ints—you may want to consider shelve.
If your structure is rigid enough, you can use key-value database for the top level, but access the values through a ctypes.Structure, or de/serialize with struct. But usually, if you can do that, you can also eliminate the top level, at which point your whole thing is just one big Structure or Array.
At that point, you can just use a plain file for storage—either mmap it (for ctypes), or just open and read it (for struct).
Or use multiprocessing's Shared ctypes Objects to access your Structure directly out of a shared memory area.
Meanwhile, if you don't actually need all of the cache data all the time, just bits and pieces every once in a while, that's exactly what databases are for. Again, anydbm, etc. may be all you need, but if you've got complex structure, draw up an ER diagram, turn it into a set of tables, and use something like MySQL.
"python in-memory data will NOT be persisted across all requests to my knowledge, unless I'm terribly mistaken."
you are mistaken.
the whole point of using uwsgi over, say, the CGI mechanism is to persist data across threads and save the overhead of initialization for each call. you must set processes = 1 in your .ini file, or, depending on how uwsgi is configured, it might launch more than 1 worker process on your behalf. log the env and look for 'wsgi.multiprocess': False and 'wsgi.multithread': True, and all uwsgi.core threads for the single worker should show the same data.
you can also see how many worker processes, and "core" threads under each, you have by using the built-in stats-server.
that's why uwsgi provides lock and unlock functions for manipulating data stores by multiple threads.
you can easily test this by adding a /status route in your app that just dumps a json representation of your global data object, and view it every so often after actions that update the store.
You said nothing about writing this data back, is it static? In this case, the solution is every simple, and I have no clue what is up with all the "it's not feasible" responses.
Uwsgi workers are always-running applications. So data absolutely gets persisted between requests. All you need to do is store stuff in a global variable, that is it. And remember it's per-worker, and workers do restart from time to time, so you need proper loading/invalidation strategies.
If the data is updated very rarely (rarely enough to restart the server when it does), you can save even more. Just create the objects during app construction. This way, they will be created exactly once, and then all the workers will fork off the master, and reuse the same data. Of course, it's copy-on-write, so if you update it, you will lose the memory benefits (same thing will happen if python decides to compact its memory during a gc run, so it's not super predictable).
I have never actually tried it myself, but could you possibly use uWSGI's SharedArea to accomplish what you're after?
I'm working on a web app with very aggressive caching. Virtually every component of the web app: views, partial views, controller output, disk loads, REST-API calls, Database queries. Everything that can be cached, at any level, is cached, all using decorators.
Naturally, this is blazing fast, since the vast majority of the HTML-generation comprises pure functions, with very few loads from disk/REST APIs. Furthermore, what few disk loads/database queries/REST API queries i perform are also cached until invalidated, so unless something just changed, they are really fast too.
So everything is blazing fast, but there is a hitch: all this stuff is being cached in memory, in one huge global dictionary in my WSGI process, and hence can be stored directly without serialization. Once i start putting stuff in memcached, the time taken for cache hits doesn't change too much, but putting stuff in cache starts taking much longer. In general that's ok, but the initial "fill cache" generation of each page goes from ~900ms (which is already pretty fast considering how many flat files it reads from disk) to about ~9000ms. For reference, generating an arbitrary page takes something like 10ms once the cache is warmed up.
Profiling the code, the vast majority of the time is going to cPickle. So the question is, how can I make this faster? Are there any in-memory caches which I can directly pass my objects to without serialization? Or some way to make caching my huge pile of objects faster? I could just go without a persistent memcached, but then my performance (or lack thereof) will be at the whim of the Apache/WSGI process manager.
If you are serializing Python objects and not simple datatypes, and have to use pickle, try cPickle.HIGHEST_PROTOCOL:
my_serialized_object = cPickle.dumps(my_object, cPickle.HIGHEST_PROTOCOL)
The default protocol is compatible with older versions of Python, but you likely don't care about that.
I just did a simple benchmark with a 1000 key dict and it was almost an order-of-magnitude faster.
UPDATE: Since you appear to already be using the highest protocol, you are going to have to do some extra work to get more performance. Here is what I would do at this point:
Identify which classes are the slowest to pickle
Create a pair of methods in the class to implement a faster serialization method, say _to_string() and _from_string(s). The actual serialization can be tailored to what the object encompasses and how it's going to be used. For example, some objects may really only contain a simple string, such as a rendered template, and some may actually be sent to a browser as JSON, in which case you can simply serialize to JSON and serve it directly. Use the timeit module to ensure that your method is actually faster
In your decorator, check hasattr(object, '_to_string') and use that instead, if it exists
This method lets you tackle the worst classes first, and introduces minimal disruption to the code base.
I am storing a table using python and I need persistence.
Essentially I am storing the table as a dictionary string to numbers. And the whole is stored with shelve
self.DB=shelve.open("%s%sMoleculeLibrary.shelve"%(directory,os.sep),writeback=True)
I use writeback to True as I found the system tends to be unstable if I don't.
After the computations the system needs to close the database, and store it back. Now the database (the table) is about 540MB, and it is taking ages. The time exploded after the table grew to about 500MB. But I need a much bigger table. In fact I need two of them.
I am probably using the wrong form of persistence. What can I do to improve performance?
For storing a large dictionary of string : number key-value pairs, I'd suggest a JSON-native storage solution such as MongoDB. It has a wonderful API for Python, Pymongo. MongoDB itself is lightweight and incredibly fast, and json objects will natively be dictionaries in Python. This means that you can use your string key as the object ID, allowing for compressed storage and quick lookup.
As an example of how easy the code would be, see the following:
d = {'string1' : 1, 'string2' : 2, 'string3' : 3}
from pymongo import Connection
conn = Connection()
db = conn['example-database']
collection = db['example-collection']
for string, num in d.items():
collection.save({'_id' : string, 'value' : num})
# testing
newD = {}
for obj in collection.find():
newD[obj['_id']] = obj['value']
print newD
# output is: {u'string2': 2, u'string3': 3, u'string1': 1}
You'd just have to convert back from unicode, which is trivial.
Based on my experience, I would recommend using SQLite3, which comes with Python. It works well with larger databases and key numbers. Millions of keys and gigabytes of data is not a problem. Shelve is totally wasted at that point. Also having separate db-process isn't beneficial, it just requires more context swaps. In my tests I found out that SQLite3 was the preferred option to use, when handling larger data sets locally. Running local database engine like mongo, mysql or postgresql doesn't provide any additional value and also were slower.
I think your problem is due to the fact that you use the writeback=True. The documentation says (emphasis is mine):
Because of Python semantics, a shelf cannot know when a mutable
persistent-dictionary entry is modified. By default modified objects
are written only when assigned to the shelf (see Example). If the
optional writeback parameter is set to True, all entries accessed are
also cached in memory, and written back on sync() and close(); this
can make it handier to mutate mutable entries in the persistent
dictionary, but, if many entries are accessed, it can consume vast
amounts of memory for the cache, and it can make the close operation
very slow since all accessed entries are written back (there is no way
to determine which accessed entries are mutable, nor which ones were
actually mutated).
You could avoid using writeback=True and make sure the data is written only once (you have to pay attention that subsequent modifications are going to be lost).
If you believe this is not the right storage option (it's difficult to say without knowing how the data is structured), I suggest sqlite3, it's integrated in python (thus very portable) and has very nice performances. It's somewhat more complicated than a simple key-value store.
See other answers for alternatives.
How much larger? What are the access patterns? What kinds of computation do you need to do on it?
Keep in mind that you are going to have some performance limits if you can't keep the table in memory no matter how you do it.
You may want to look at going to SQLAlchemy, or directly using something like bsddb, but both of those will sacrifice simplicity of code. However, with SQL you may be able to offload some of the work to the database layer depending on the workload.
I have a python module that makes use of a huge dictionary global variable, currently I put the computation code in the top section, every first time import or reload of the module takes more then one minute which is totally unacceptable. How can I save the computation result somewhere so that the next import/reload doesn't have to compute it? I tried cPickle, but loading the dictionary variable from a file(1.3M) takes approximately the same time as computation.
To give more information about my problem,
FD = FreqDist(word for word in brown.words()) # this line of code takes 1 min
Just to clarify: the code in the body of a module is not executed every time the module is imported - it is run only once, after which future imports find the already created module, rather than recreating it. Take a look at sys.modules to see the list of cached modules.
However, if your problem is the time it takes for the first import after the program is run, you'll probably need to use some other method than a python dict. Probably best would be to use an on-disk form, for instance a sqlite database, one of the dbm modules.
For a minimal change in your interface, the shelve module may be your best option - this puts a pretty transparent interface between the dbm modules that makes them act like an arbitrary python dict, allowing any picklable value to be stored. Here's an example:
# Create dict with a million items:
import shelve
d = shelve.open('path/to/my_persistant_dict')
d.update(('key%d' % x, x) for x in xrange(1000000))
d.close()
Then in the next process, use it. There should be no large delay, as lookups are only performed for the key requested on the on-disk form, so everything doesn't have to get loaded into memory:
>>> d = shelve.open('path/to/my_persistant_dict')
>>> print d['key99999']
99999
It's a bit slower than a real dict, and it will still take a long time to load if you do something that requires all the keys (eg. try to print it), but may solve your problem.
Calculate your global var on the first use.
class Proxy:
#property
def global_name(self):
# calculate your global var here, enable cache if needed
...
_proxy_object = Proxy()
GLOBAL_NAME = _proxy_object.global_name
Or better yet, access necessery data via special data object.
class Data:
GLOBAL_NAME = property(...)
data = Data()
Example:
from some_module import data
print(data.GLOBAL_NAME)
See Django settings.
I assume you've pasted the dict literal into the source, and that's what's taking a minute? I don't know how to get around that, but you could probably avoid instantiating this dict upon import... You could lazily-instantiate it the first time it's actually used.
You could try using the marshal module instead of the c?Pickle one; it could be faster. This module is used by python to store values in a binary format. Note especially the following paragraph, to see if marshal fits your needs:
Not all Python object types are supported; in general, only objects whose value is independent from a particular invocation of Python can be written and read by this module. The following types are supported: None, integers, long integers, floating point numbers, strings, Unicode objects, tuples, lists, sets, dictionaries, and code objects, where it should be understood that tuples, lists and dictionaries are only supported as long as the values contained therein are themselves supported; and recursive lists and dictionaries should not be written (they will cause infinite loops).
Just to be on the safe side, before unmarshalling the dict, make sure that the Python version that unmarshals the dict is the same as the one that did the marshal, since there are no guarantees for backwards compatibility.
If the 'shelve' solution turns out to be too slow or fiddly, there are other possibilities:
shove
Durus
ZopeDB
pyTables
shelve gets really slow with large data sets. I've been using redis quite successfully, and wrote a FreqDist wrapper around it. It's very fast, and can be accessed concurrently.
You can use a shelve to store your data on disc instead of loading the whole data into memory. So startup time will be very fast, but the trade-off will be slower access time.
Shelve will pickle the dict values too, but will do the (un)pickle not at startup for all the items, but only at access time for each item itself.
A couple of things that will help speed up imports:
You might try running python using the -OO flag when running python. This will do some optimizations that will reduce import time of modules.
Is there any reason why you couldn't break the dictionary up into smaller dictionaries in separate modules that can be loaded more quickly?
As a last resort, you could do the calculations asynchronously so that they won't delay your program until it needs the results. Or maybe even put the dictionary in a separate process and pass data back and forth using IPC if you want to take advantage of multi-core architectures.
With that said, I agree that you shouldn't be experiencing any delay in importing modules after the first time you import it. Here are a couple of other general thoughts:
Are you importing the module within a function? If so, this can lead to performance problems since it has to check and see if the module is loaded every time it hits the import statement.
Is your program multi-threaded? I have seen occassions where executing code upon module import in a multi-threaded app can cause some wonkiness and application instability (most notably with the cgitb module).
If this is a global variable, be aware that global variable lookup times can be significantly longer than local variable lookup times. In this case, you can achieve a significant performance improvement by binding the dictionary to a local variable if you're using it multiple times in the same context.
With that said, it's a tad bit difficult to give you any specific advice without a little bit more context. More specifically, where are you importing it? And what are the computations?
Factor the computationally intensive part into a separate module. Then at least on reload, you won't have to wait.
Try dumping the data structure using protocol 2. The command to try would be cPickle.dump(FD, protocol=2). From the docstring for cPickle.Pickler:
Protocol 0 is the
only protocol that can be written to a file opened in text
mode and read back successfully. When using a protocol higher
than 0, make sure the file is opened in binary mode, both when
pickling and unpickling.
I'm going through this same issue...
shelve, databases, etc... are all too slow for this type of problem. You'll need to take the hit once, insert it into an inmemory key/val store like Redis. It will just live there in memory (warning it could use up a good amount of memory so you may want a dedicated box). You'll never have to reload it and you'll just get looking in memory for keys
r = Redis()
r.set(key, word)
word = r.get(key)
Expanding on the delayed-calculation idea, why not turn the dict into a class that supplies (and caches) elements as necessary?
You might also use psyco to speed up overall execution...
OR you could just use a database for storing the values in? Check out SQLObject, which makes it very easy to store stuff to a database.
There's another pretty obvious solution for this problem. When code is reloaded the original scope is still available.
So... doing something like this will make sure this code is executed only once.
try:
FD
except NameError:
FD = FreqDist(word for word in brown.words())