High performance and actively maintained server-side session framework for Pyramid and Redis.
Project description
IMPORTANT
pyramid_session_redis
is an actively maintained fork of pyramid_redis_sessions
(`ericrasmussen/pyramid_redis_sessions``), with many improvements and API changes designed for high performance (particularly with servers under load) and a slightly different API for developer convenience.
This package is following a multi-version release process.
The 1.5.x branch is a stable API and designed for Pyramid < 2. This is the last branch that will support Pyramid 1.x. Projects using this package on Pyramid 1.x should peg their requirements to the 1.5 branch
"pyramid_session_redis>1.5,<1.6"
The 1.6 branch will be required for Pyramid2 due to breaking changes.
Prior branches:
The 1.2.x branch is in maintenance mode as of 1.2.2, and will culminate in a final 1.3.0 release.
The 1.4.x branch is under active development and subject to change. It culminated in a stable 1.5.0 API release.
The 1.2.x branch and earlier are largely compatible with pyramid_redis_sessions
as-is. If you are using this, you should pin your installs to <=1.3.0
or <1.3
.
The 1.4.x branch and later have several design changes and are not a drop-in replacement. Some kwargs may have changed. The structure of the package has changed as well, and advanced users who leverage the internals will need to upgrade. The package remains a plug-and-play Pyramid sessions interface.
IMPORTANT: The internal payload structure has changed in the 1.4 branch, and is no longer compatible with existing 1.2 sessions (they will be invalid). I am open to PRs that can handle a graceful fallback.
The 1.2 format guaranteed sessions to be in this format:
{'managed_dict': {},
'created': INT,
'timeout': INT,
}
The 1.4 version streamlines the keys for a lighter footprint and generally looks like this:
{'m': {}, # managed_dict
'c': INT, # created
'v': SESSION_API_VERSION, # session INT
# the following are optional and not guaranteed to be around
't': INT, # timeout, not gua
'x': INT, # expiry
}
- keys were shortened to 1 letter
- an api version is now used, to handle graceful changes in the future
- the timeout is no longer guaranteed to be present (it may now be handled on redis)
- an expiry timeout may also exist
Key Differences:
Depending on your needs, this package is probably more desirable than the original project. This package significantly cuts down on the communication between Redis and Pyramid vs the original implementation. Some options are offered to minimize the size of payloads as well.
This package contains a lot of hooks and features to aid developers who are using this in high-traffic situations. This package does not recommend a "best deployment", but offers different strategies for creating a best deployment under different circumstances.
Through 1.2.x
- The original package communicates with Redis on most attribute access and writes. The traffic can be too much on some implementations.
pyramid_session_redis
will queue a singlepersist
orrefresh
task using Pyramid'sadd_finished_callback
hook. - The original version used
EXISTS
to check if a session existed or not, then proceeded toGET
orSET
a new session.pyramid_session_redis
will immediately attempt aGET
, and willSET
a new session on failure. This eliminates a call. - Separate calls to
SET
andEXPIRE
were replaced with a singleSETEX
- A flag can be set to enable a LRU Cache (least recently used) mode. No expiry data will be sent to Redis, allowing the Redis server to handle the LRU logic itself
- The active
session
is decoupled from the request attribute (ie, this can handle a "session" set up on alternate attributes) - The original library does not detect changes in nested dictionaries. This package uses
hashlib.md5
to fingerprint the serialized value on read; if no changes were detected a failsafe will serialize+md5 the data to decide if a write should occur. This can be disabled by settingdetect_changes
to False. - The original raises a fatal error if a session can not be deserialized. by passing in
deserialized_fails_new
to the constructor, you can create a new session on deserialization errors. - Support for disabling sessions on CDN generated content via
func_check_response_allow_cookies
- Thanks to @ github/hongyuan1306, token generation has been consolidated to use python3's stdlib (or reimplemented if not available). tokens are also 32, not 20, chars.
- redis is supported in a LRU mode (see http://redis.io/topics/lru-cache) by setting the option
set_redis_ttl
toFalse
(by default, it isTrue
). This will eliminate calls toEXPIRE
and will useSET
instead ofSETEX
. - in the 1.2.x branch the created time can be set to an integer via
use_int_time=True
. This will cast thecreated
time via "int(math.ceil(time.time()))". This can reduce a payload by several bits.
Other Updates 1.4.x+
- only int() time is supported.
- sessions now have version control to support future upgrades via a "version"
v
key. - the format of the internal payload was rewritten, the encoded payload now uses 1-letter keys instead of words. this should offset the addition of an expires timestamp and version id.
- there was no logic for python timeout control (whoops!) this has been fixed. an "expires"
x
key now tracks the expiration. - added a
timeout_trigger
option. this will defer expiry data updates to lower usage on Redis. This is explained below in more detail. - In high load situations, Redis can have performance and storage issues because in the original package sessionIDs are created on every request (such as a getting spidered by a botnet that does not respect sessions). in this package, a 'lazycreate' method is used. a session_id/cookie will not be generated unless a session is needed in the callback routine. in order to generate session_id/cookie beforehand, one can use the
RedisSession.ensure_id
function. To safely check if a session_id exists, one can use theRedisSession.session_id_safecheck
method as well. - added
func_invalid_logger
to constructor. this can be used to log invalid sessions. it is incredibly useful when integrated with a statsd system. (see below) - added
set_redis_ttl_readheavy
to the factory and session constructors. This option will optimize sessions which have Redis-maintained TTLs by executing a GET+EXPIRE together within a pipeline.
Notes:
If set_redis_ttl
is False, it does not imply there is no timeout at all -- only that Redis will not be sent timeout data via SETEX
or EXPIRE
. Timeout data will still be stored in Python.
If Redis is functioning as an LRU Cache, abandoned sessions will never be seen by Python, but will eventually be cleared out to make room for new sessions by the inherent Redis LRU logic.
Timeout data stored in Python is relatively small when compared to the timeout data stored in Redis.
If you want to NEVER have sessions timeout, set the initial timeout
to "0" or "None".
Setting a timeout_trigger
will require Python to track the expiry.
Enableing set_redis_ttl_readheavy
requires a timeout
and set_redis_ttl
; it also requires not enabling timeout_trigger
or python_expires
.
Examples:
Timeout in Python, with Redis TTL via SETEX
/EXPIRE
:
timeout = 60
Timeout in Python, no Redis TTL (only SET
used)
timeout = 60
assume_redis_ttl = True
No Timeout in Python, no Redis TTL (only SET
used)
timeout = 0 # or None
assume_redis_ttl = True
Warnings:
This package does not incorporate any sort of locking for session updating or handling of race conditions.
This should not be a problem for the vast majority of users, and is a feature not present on any known Pyramid session providers.
For more information, please see:
- https://github.com/jvanasco/pyramid_session_redis/issues/9
- https://github.com/Pylons/pyramid/issues/3041
Timeout Triggers
A timeout trigger can be used to limit the amount of updates/writes. It may be more beneficial to your usage pattern.
For the sake of clarity, I'll use an oversimplification that Redis essentially has two different internal data stores that are used independently: one for a Key's payload and another for the Key's expiry.
In the 'classic' behavior (project this was forked from): every time you access an existing session, the GET is followed by sending Redis a new EXPIRE time - essentially every "read" has a corresponding "write" for the Redis-tracked expiry record.
In order to minimize the writes via SETEX, I introduced the timeout trigger. The trigger works by storing some timeout information in the Redis data payload, and using that information to determine when to send a write. Instead of having a GET+EXPIRE for every read, we only have a single GET and eliminate the write caused by the EXPIRE. This has a maintenance cost though - once we hit the timeout trigger, instead of just the GET we need to update the internal timeout payload and issue a SET.
Going back to your situation: when the user stops activity at 40 minutes in, if the timeout trigger is enabled then there has never been an update to the internal payload or Redis about the user activity since the session was first created. The purpose of the trigger is to defer that "write" operation.
In order to make a session valid for "reading" for around an hour, you'd want to do something like:
- a two-hour session with a 10 minute trigger, or
- a one-hour session with a 50 minute trigger
You can also disable the functionality by setting the trigger to 0. Up to a few thousand daily users, you shouldn't have any issues with the overhead of the "classic" mode. When you hit 10k users and/or start to have clustered web servers communicating with dedicated Redis instances, setting a new EXPIRE after every read operation becomes something you want to avoid.
Scenario 1 - Classic Redis
In the typical "classic" Redis usage pattern, the session usage is refreshed via an EXPIRE
call on every session view
This is useful, but means many session operations will trigger two Redis calls (GET
+ EXPIRE
). On a high performance system, this can be a lot.
This is a typical scenario with refreshing:
timeout = 200
time | Redis calls | timeout |
---|---|---|
0 | GET +SETEX |
200 |
100 | GET +EXPIRE |
300 |
200 | GET +EXPIRE |
400 |
300 | GET +EXPIRE |
500 |
400 | GET +EXPIRE |
600 |
500 | GET +EXPIRE |
700 |
Scenario 2 - Timeout Trigger
The 1.4.x branch introduces a timeout_trigger
to augment the session's timeout
.
Whereas a timeout
states how long a session is good for, a timeout_trigger
defers how long a session's refresh should be deferred for:
Given the following example, the package will use a 1200s timeout for requests, but only trigger an update of the expiry time when the current time is within 600s of the expiry
timeout = 1200
timeout_trigger = 600
The following timeline would occur
time | Redis calls | timeout | next threshold |
---|---|---|---|
0 | GET +SET * |
1200 | 600 |
1 | GET |
1200 | 600 |
... | |||
599 | GET |
1200 | 600 |
600 | GET +SET * |
1800 | 1200 |
601 | GET |
1800 | 1200 |
... | |||
1199 | GET |
1800 | 1200 |
1200 | GET +SET * |
2400 | 1800 |
- This method is compatible with setting a TTL in redis via
SETEX
or doing everything within Python if redis is in a LRU mode
The removes all calls to EXPIRE
before the threshold is reached, which can be a considerable savings in read-heavy situations
The caveat to this method: an expiry timestamp must be stored within the payload AND updating the timeout requires a SET
operation.
set_redis_ttl_readheavy
This is a new option in 1.4.2
which should improve performance on readheavy installations BUT may degrade performance on writeheavy installations. This option will aggregate a GET+EXPIRE on every read.
set_redis_ttl_readheavy
requires the following:
- a
timeout
value is set set_redis_ttl
isTrue
timeout_trigger
is NOT setpython_expires
is NOT True
The default behavior of this library during a read-only request is this:
- On session access, query Redis via
redis.GET {session_id}
- In a Pyramid callback, update Redis via
redis.EXPIRE {session_id} {expiry}
During a read-write:
- On session access, query Redis via
redis.GET {session_id}
- In a Pyramid callback, update Redis via
redis.SETEX {session_id} {payload} {expiry}
The new set_redis_ttl_readheavy
changes this. If enabled, during a read-only request the behavior will be lighter on the Redis instance:
- On session access, open a pipeline with two Redis commands:
pipeline.GET {session_id}
,pipeline.EXPIRE {session_id} {expiry}
. - In a Pyramid callback, the duplicate expire is suppressed.
However during a read-write, the activity will be:
- On session access, open a pipeline with two Redis commands:
pipeline.GET {session_id}
,pipeline.EXPIRE {session_id} {expiry}
. - In a Pyramid callback, update Redis via
redis.SETEX {session_id} {payload} {expiry}
Read-heavy applications should see a slight performance bump via the piplined GET+EXPIRE, however write-heavy applications are likely to see a performance degradation as it adds an extra EXPIRE to every request.
Invalid Logging
The default behavior of this library is to silently create new session when bad session data is encountered, such as a cookie with an invalid id or corrupted datastore. A graceful "new session" is the ideal situation for end-users.
The problem with that strategy is that problems in code or your application stack can be hidden, and you might not know about a bad datastore.
The 1.4 release introduces func_invalid_logger
to the factory constructor.
This can be used to track the invalid sessions that are safely caught and silently upgraded
How? The package tracks why a session is invalid with variant classes of pyramid_session_redis.exceptions.InvalidSession
Specifically there are the following classes:
InvalidSession(Exception)
Catchall base classInvalidSession_NoSessionCookie(InvalidSession)
The session is invalid because there is no cookie. This is the same as "new session".InvalidSession_NotInBackend(InvalidSession)
The session id was not in the backendInvalidSession_DeserializationError(InvalidSession)
Error deserializing. This is raised ifdeserialized_fails_new
is True. Otherwise the exception is wrapped in aRawDeserializationError
and raised without being caught.InvalidSession_PayloadTimeout(InvalidSession)
The inner python payload timed outInvalidSession_PayloadLegacy(InvalidSession)
The session is running on an earlier version
The factory accepts a func_invalid_logger
callable argument. The input is the raised exception BEFORE a new cookie is generated, and will be the request and an instance of InvalidSession
.
from pyramid_session_redis.exceptions import *
from my_statsd import new_statsd_client()
statsd_client = new_statsd_client()
def my_logger(request, raised_exception):
"""
raised_exception will be an instance of InvalidSession
log the exception to statsd for metrics
"""
if isinstance(raised_exception, InvalidSession_NoSessionCookie):
statsd_client.incr('invalid_session.NoSessionCookie')
elif isinstance(raised_exception, InvalidSession_NotInBackend):
statsd_client.incr('invalid_session.NotInBackend')
elif isinstance(raised_exception, InvalidSession_DeserializationError):
statsd_client.incr('invalid_session.DeserializationError')
factory = RedisSessionFactory(...
func_invalid_logger=my_logger,
...
)
The func_invalid_logger
argument may be provided as a dotted-notation string in a settings file.
Uncaught Errors
The exception pyramid_session_redis.exceptions.RawDeserializationError
will be raised if deserialization of a payload fails and deserialized_fails_new
is not True
. The first arg will be the caught exception. This allows for a standard error across multiple deserialization options.
FAQ:
Q. What serialization is used?
A. Serialization is used at two points
-
"serializing" the session_id to encode into a signed cookie. originally this was handled by
pickle
(de)serialization (<= v1.5.0), but this step been removed (v1.5.1) as the default id generation creates a cookie-safe value. -
serializing the session data. this is handled by pickle. session data is created on the server by your application and stays on the server, so
pickle
is safe to use. you can override this with a custom JSON, msgpack or other (de)serialization routine.
Further Reading:
For more information about Redis performance under python please see an associated project:
Until Nov 2016 this was maintained as jvanasco/pyramid_redis_sessions
- The master branch for
jvanasco/pyramid_redis_sessions
is "custom_deployment" - The branched named "master" is the upstream source from ericrasmussen
As of Nov 2016, this was forked into it's own project to allow for distribution.
All support is handled via GitHub : https://github.com/jvanasco/pyramid_session_redis
ToDo
see TODO.md
Overview
pyramid_redis_sessions
is a server-side session library for the Pyramid Web
Application Development Framework, using Redis for storage. This library
implements the Pyramid ISession interface <http://docs.pylonsproject.org/projects/pyramid/en/latest/api/interfaces.html#pyramid.interfaces.ISession>
_.
Why Use Redis for Your Sessions
Redis is fast, widely deployed, and stable. It works best when your data can
fit in memory, but is configurable and still quite fast when you need to sync
to disk. There are plenty of existing benchmarks, opinion pieces, and articles
if you want to learn about its use cases. But for pyramid_redis_sessions
, I'm
interested in it specifically for these reasons:
- it really is bleeping fast (choose your own expletive)
- it has a very handy built-in mechanism for setting expirations on keys
- the watch mechanism is a nice, lightweight alternative to full transactions
- session data tends to be important but not mission critical, but if it is...
- it has configurable
persistence <http://redis.io/topics/persistence>
_
Documentation
To get up and running as fast as possible, check out the
Getting Started <http://pyramid-redis-sessions.readthedocs.org/en/latest/gettingstarted.html>
_
guide.
You can also read the
full documentation <http://pyramid-redis-sessions.readthedocs.org/en/latest/index.html>
_
on Read the Docs.
Support
You can report bugs or open feature/support requests via GitHub : https://github.com/jvanasco/pyramid_session_redis
License
pyramid_session_redis
is available under a FreeBSD-derived license. See
LICENSE.txt <https://github.com/jvanasco/pyramid_session_redis/blob/master/LICENSE.txt>
_
for details.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Hashes for pyramid_session_redis-1.5.3.tar.gz
Algorithm | Hash digest | |
---|---|---|
SHA256 | c40a6fb735ec15c27dce12ebd1d6355509e527f1958bddc991cb09d1c5254779 |
|
MD5 | 5d84c4198a14c8d1154e52efc89d34b8 |
|
BLAKE2b-256 | 3f45426e797904cf0553e8b48b719dbbd37161c83e22379a1819d23717652773 |