sqlalchemy-collectd

Send statistics on SQLAlchemy connection and transaction metrics used by Python applications to the collectd service.

sqlalchemy-collectd works as a SQLAlchemy plugin invoked via the database URL, so can be used in any SQLAlchemy application (1.1 or greater) that accepts arbitrary connection URLs. The plugin is loaded using setuptools entrypoints and no code changes to the application are required. There are no dependencies on database backends or drivers.

sqlalchemy-collectd is oriented towards providing a unified view of application-side database metrics in sprawling, many-host / many-process environments that may make use of any number of topologically complicating technologies such as database clusters, proxy servers, large numbers of client applications, multi-process applications, and containers.

What's collectd?

collectd is a statistics collection daemon that is easy to run. It serves as a collector and re-broadcaster of runtime statistics for a wide variety of performance and other metrics. Once a set of stats are in collectd, they can be broadcast out virtually anywhere, including to RRD databases and front-ends, to metrics reporting applications like Graphite and Grafana, and to other collectd servers.

Architecture Nutshell

sqlalchemy-collectd gathers its statistics from within the Python application itself, and delivers live metrics to a collectd service over UDP. To achieve this, it's client portion is loaded within the process as a SQLAlchemy engine plugin that attaches listeners to a sqlalchemy.engine.Engine object as well as the connection pool within it. A background thread running within each process periodically sends a snapshot of statistics out over UDP.

On the collectd side, a Python plugin listens on the same UDP port and aggregates statistics received from any number of Python processes and hosts, and then delivers them to the collectd engine itself as a series of per-host and per-program statistics.

A key goal of this architecture is to allow a Python program that uses multiple processes (e.g. via Python multiprocessing or just plain fork) to report unified information on each engine/connection pool within each subprocess, as well as to track multiple instances of the same application running from many hosts (and of course it can track any number of applications and hosts simultaneously). Having a full client/server model allows the collectd service itself to be located not only on the same host as the application itself, but on any other host on the network.

The network model itself makes use of collectd's own binary protocol; while not strictly necessary, this is because originally the plan was to use the collectd "network" plugin as the receiver, however after observing limitations in collectd's "aggregation" plugin this was replaced with a full Python plugin that does everything needed in a much more straightforward way.

How is this different from using database monitoring?

When you run a database like MySQL or Postgresql, there's lots of ways to see activity in the database; you can list out statistics regarding connections, transactions, threads / processes in use, and in most cases you can integrate these server-side statistics with collectd itself to watch trends in real time.

However, while gathering stats from the server can provide insights into activity, including being able to look at the originating host as well as the specific database(s) being accessed by each client, in a large scale environment it's difficult to get a unified, real-time picture for how each process on each host is making use of its database connections, especially if there are additional layers of indirection between application and databases present, such as proxy servers like HAProxy, ProxySQL or PGBouncer, as well as when databases and/or applications are containerized and potentially hopping over additional network translation layers. This kind of analysis requires being able to relate database connections reported by the database to the originating hosts and individual processes on each host.

SQLAlchemy-oriented applications usually make use of process-local connection pools as well, and it is valuable to be able to see how well these pools are being utilized, which includes being able to see how many connections are sitting idle vs. how often does the application need to create new connections in order to respond to requests. These are still things that can probably be gleaned from the database itself from things like connection idle time, but especially when layers of indirection are in place, it's simpler to get the performance metrics you care about (e.g., how well are the applications performing) from the applications themselves, as they can give you the exact information about what they are doing without having to reverse-engineer it from database servers and network status.

Of course, this assumes the applications are Python applications using SQLAlchemy. Which of course they should be! :)

Installation

To use SQLAlchemy-collectd, you need to have:

• SQLAlchemy-collectd installed as a Python library alongside SQLAlchemy itself, in all Python environments that run a SQLAlchemy-oriented application.
• The collectd service itself somewhere on the network.
• The collectd-python plugin, which may be delivered as a separate package depending on distribution
• SQLAlchemy-collectd installed as a Python library alongside the collectd server itself, either as part of the system Python which collectd-python accesses by default, or the SQLAlchemy-collectd application can be extracted into any arbitrary location that can be set up as an additional system path with collectd.

Without using a package manager, SQLAlchemy-collectd can be installed via pip using:

pip install sqlalchemy-collectd

Configuration

Configuration involves both a client-side configuration as well as a server side configuration. Both are very simple.

Client

SQLAlchemy applications use a database connection URL, usually loaded from a configuration system of some kind. Wherever this URL is for your target application, basically add ?plugin=collectd to it (or &plugin=collectd if other query parameters already there). Such as:

mysql+pymysql://user:password@databasehost/dbname?charset=utf8&plugin=collectd

Using a URL as above, the sqlalchemy-collectd plugin will be pulled in where it will attempt to send messages to a collectd server listening on localhost port 25827 (note this is one higher than the default collectd network plugin port of 25826).

Destination Host

To send stats to collectd on a different host, add collectd_host (currently ipv4 only) and optionally collectd_port:

mysql+pymysql://user:password@databasehost/dbname?charset=utf8&plugin=collectd&collectd_host=172.18.5.2&collectd_port=25827

Program Name

Another important configuration is the "program name" - this is the application name that sqlalchemy-collectd will report within statistics. Within a particular "program name" on a particular host, statistics are aggregated across all processes, regardless of parent process.

By default, the "program name" comes from sys.argv[0], but this is not always what's desired; for example, if you're running from within mod_wsgi, this will likely return httpd which is more vague that most would prefer. Additionally, a single application might create connections to multiple databases for different purposes, and one might want to separate the reporting for these into different sections. To set up this program name, add collectd_program_name:

mysql+pymysql://user:password@databasehost/dbname?charset=utf8&plugin=collectd&collectd_program_name=nova_api&collectd_host=172.18.5.2

With the above URL, all Python processes that use this URL on a single host will aggregate their connection use statistics under the name nova_api.

Startup

After the URL is configured, the vast majority of applications probably need to be restarted for the change to take effect.

The plugin will transparently spawn a background thread for each individual process that starts up which also connects to the database (don't worry, these work if you are using gevent, eventlet, asyncio, gunicorn, etc. threads are your friend).

Server

sqlalchemy-collectd uses a Python plugin, so in your collectd.conf or in a collectd.d/sqlalchemy.conf file, assuming a system-installed sqlalchemy-collectd:

LoadPlugin python
<Plugin python>
    LogTraces true

    Import "sqlalchemy_collectd.server.plugin"

    <Module "sqlalchemy_collectd.server.plugin">
        # ipv4 only for the moment
        listen "0.0.0.0" 25827

        # set to "debug" to show messages received
        loglevel "info"

    </Module>
</Plugin>

Above, the plugin will listen for UDP on port 25827 of the default network interface. It can also be configured to listen on "localhost" or any other IP number (currently ipv4 only) on the host.

Custom Module Path

To reference sqlalchemy-collectd extracted into an arbitrary file location, add ModulePath:

LoadPlugin python
<Plugin python>
        ModulePath "/path/to/sqlalchemy-collectd/"
    LogTraces true

    Import "sqlalchemy_collectd.server.plugin"

    <Module "sqlalchemy_collectd.server.plugin">
        # ipv4 only for the moment
        listen "0.0.0.0" 25827

        # set to "debug" to show messages received
        loglevel "info"
    </Module>
</Plugin>

For further information about the Python plugin system see collectd-python.

The collectd server is typically restarted for the configurational change to take effect.

TODO

• ipv6 support
• security layer (e.g. network packet signing / encryption)

Stats

Now that sqlalchemy-collectd is running, what stats can we see?

Supposing we have the plugin turned on for the applications neutron and nova, the namespace we would see in a tool like graphana would look like:

hostname
        sqlalchemy-host
                count-checkedin
                count-checkedout
                count-connections
                count-detached
                count-numpools
                count-numprocs
                derive-checkouts
                derive-connects
                derive-disconnects
                derive-invalidated
                derive-commits
                derive-rollbacks
                derive-transactions

        sqlalchemy-neutron
                count-checkedin
                count-checkedout
                count-connections
                count-detached
                ... everything else

        sqlalchemy-nova
                count-checkedin
                count-checkedout
                count-connections
                count-detached
                ... everything else

Above, we first see that all stats are grouped per-hostname. Within that, we have a fixed plugin instance called "host", which renders as sqlalchemy-host. This represents aggregated statistics for the entire host, that is, statistics that take into account all database connections used by all applications (that use sqlalchemy-collectd) on this particular host.

Following that, we can see there are groups for the individual program_name we set up, for nova and neutron we get stats aggregated for that name specifically.

The statistics themselves are labeled count-<name> or derive-<name>, which correspond to pre-supplied collectd types count and derive (see "collectd types" below for why the naming is done this way). The stats labeled count are integers representing the current count of a resource or activity:

• count-checkedin - current number of connections that are checked in to the connection pool

• count-checkedout - current number of connections that are checked out from the connection pool, e.g. are in use by the application to talk to the database.

• count-connections - total number of connections to the database at this moment, checked out, checked in, detached, or soft-invalidated.

• count-detached - total number of connections that are detached; meaning they have been disconnected from the engine/pool using the .detach() method but are still being used as a database connection.

• count-numpools - the number of connection pools in use. A SQLAlchemy Engine features exactly one connection pool. If an application connects to two different database URLs in a process and creates two different Engine objects, then you'd have two pools. If that same application spawns off into ten subprocesses, then you have 20 or 22 pools in use, depending on how the parent uses the database also. Use count-numpools to make sure this number is what you expect. A poorly written application that is spawning a brand new Engine for each request will have a dramatically larger number here (as well as one that is changing constantly) and that is an immediate red flag that the application should be fixed.

• count-numprocs - the total number of Python processes, e.g. parent and subprocesses, that are contributing to the connection statistics in this group. This number will match count-numpools if you have one Engine per process.

  Both the count-numpools and count-numprocs values provide context to when one looks at the total connections and checkouts. If connection pools are configured to allow at most 20 connections max, and you have 10 connection pools on the host, now you can have 200 connections max to your database.

The stats labeled derive are floating point values representing a rate of activity. sqlalchemy-collectd sends these numbers to the collectd server as a total number of events occurred as of a specific timestamp; collectd then compares this to the previous value to determine the rate. How the rate is reported (e.g. number per second, etc.) depends on the reporting tools being used.

• derive-checkouts - rate of connections being checked out.
• derive-connects - rate of new connections made to the database
• derive-disconnects - rate of database connections being closed
• derive-invalidated - rate of connections that are explicitly invalidated, e.g. have encountered a connectivity error which made the program invalidate the connection. The application may or may not have tried to connect again immediately depending on how it is using this feature. See the section on "invalidated connections" below for details on this.
• derive-commits - (TODO: not implemented yet) rate of calls to transaction.commit(). This value can be used to estimate TPS, e.g. transactions per second, however note that this is limited to SQLAlchemy-explicit transactions where the Engine-level begin() / commit() methods are being invoked. When using the SQLAlchemy ORM with the Session, this rate should be tracking the rate of calls to Session.commit().
• derive-rollbacks - (TODO: not implemented yet) rate of calls to transaction.rollback().
• derive-transactions - (TODO: not implemented yet) rate of transactions overall. This should add up to the commit and rollback rates combined, however may be higher than that if the application also discards transactions and/or Session objects without calling .commit() or .rollback().

Invalidated Connections

The derive-invalidated stat records the rate of invalidations.

By invalidated, we mean the .invalidated() method on the connection is called, which marks this connection as no longer usable and marks it for refresh on next use (soft invalidation) or more commonly closes it immediately (hard invalidation). Typically, when a connection is invalidated, the application is either pre-pinging the database and will try to connect again, or it was in the middle of an operation when the database got cut off, in which case depending on how the application was designed it may or may not try the operation again.

Invalidation usually corresponds to a connection that reported a problem in being able to communicate with the database, and for which an error was raised. For this reason, the "invalidated" rate should be considered to be roughly an "error" rate - each count here usually corresponds to a connectivity error encountered by the application to which it responded by invalidating the connection, which results either in immediate or eventual reconnection.

For most invalidation scenarios, the entire pool of connections is invalidated at once using a "freshness" timestamp; any connection older than this timestamp is refreshed on next use. This is to suit the case of assuming that the database was probably restarted, so all connections need to be reconnected. These connections which have been implicitly invalidated are not included in this count.

Collectd Types

These funny names count- and derive- are an artifact of how collectd provides types. collectd has a fixed list of "types" which it lists in a file called types.db. The server does not accept type names that are not either in this file or in a separately configured custom types file, as each type is accompanied by a template for what kinds of values it carries. Annoyingly, collectd does not let us add these names within the regular .conf file, which would make it very easy for us to include our own custom names; it instead requires they be listed in completely separate file that must be explicitly referred to by absolute path within a conf file, and then to make matters worse when this option is used, we have to uncomment the location of the default types.db file in the central collectd.conf else it will no longer be able to find it. Given the choice between "very nice names" and "no need to set up three separate config files", we chose the latter :)

connmon mode

As an added feature, the connmon UX has now been integrated into SQLAlchemy-collectd. This is a console application that displays a "top"-like display of the current status of connections.

Using the configuration above, we can add a a plugin configuration for the connmon server plugin:

LoadPlugin python
<Plugin python>
    LogTraces true

    Import "sqlalchemy_collectd.server.plugin"

    <Module "sqlalchemy_collectd.server.plugin">
        # ipv4 only for the moment
        listen "0.0.0.0" 25827

        # set to "debug" to show messages received
        loglevel "info"

    </Module>

    Import "sqlalchemy_collectd.connmon.plugin"
    <Module "sqlalchemy_collectd.connmon.plugin">
        monitor "localhost" 25828

        # set to "debug" to show messaging
        #loglevel "debug"
    </Module>

</Plugin>

We can now run "connmon" on localhost port 25828:

connmon --port 25828

The connmon plugin and command line tool as of version 0.6 works independently of the "server" plugin, and may be configured by itself without the server plugin being present. It now consumes sqlalchemy-collectd events not only from the local server plugin if present, but from also from any other sqlalchemy-collectd messages that are forwarded to the server from elsewhere, typically via the "network" plugin.

Screenshot of connmon:

The SQLAlchemy Project

SQLAlchemy-collectd is part of the SQLAlchemy Project and adheres to the same standards and conventions as the core project.

Development / Bug reporting / Pull requests

Please refer to the SQLAlchemy Community Guide for guidelines on coding and participating in this project.

Code of Conduct

Above all, SQLAlchemy places great emphasis on polite, thoughtful, and constructive communication between users and developers. Please see our current Code of Conduct at Code of Conduct.

License

SQLAlchemy-collectd is distributed under the MIT license.