Polyfill-Pinba

Pure-php reimplementation of the "PHP extension for Pinba".

See http://pinba.org for the original.

Requirements

PHP 5.3 or any later version.

A Pinba server to send the data to. Known servers include http://pinba.org/ and https://github.com/badoo/pinba2. Both are available as Docker Container images for trying out.

Installation

composer require gggeek/polyfill-pinba

then set up configuration settings pinba.enabled and pinba.server in php.ini as described at https://github.com/tony2001/pinba_engine/wiki/PHP-extension#INI_Directives

That's all.

NB gathering of metrics and sending them to the server is disabled by default. You have to enable it via ini settings, unless you have added explicit calls to the pinba api in your php code.

Usage

See the API described at https://github.com/tony2001/pinba_engine/wiki/PHP-extension

A trivial usage example can be found in doc/sample.php.

For viewing the gathered metrics, check out https://github.com/intaro/pinboard, https://github.com/pinba-server/pinba-server or https://github.com/ClickHouse-Ninja/Proton

Extensions to the original API

Pinba::ini_set

If the pinba php extension is not enabled in your setup (which is most likely the case, as otherwise you would not be using this package), it is not possible from php code to modify the values for ini options pinba.enabled and pinba.server. While it is possible to set their value in php.ini, if you want to modify their value at runtime you will have instead to use methods \PinbaPhp\Polyfill\pinba::ini_set($option, $value). You should also use corresponding method \PinbaPhp\Polyfill\pinba::ini_get($option) to check it.

ini option pinba.inhibited

In case you want to keep your code instrumented with pinba_timer_add, pinba_timer_stop and similar calls but are not collecting the pinba data anymore, and you want to reduce as much as possible the overhead imposed by this package, please set in pinba.inhibited=1 php.ini.

Using pinba.enabled=0 or pinba.auto_flush=0 is not recommended in that scenario, as, while they both disable the sending of data, they do not prevent timers to be actually created.

Compatibility

We strive to implement the same API as Pinba extension ver. 1.1.2.

As for the server side, the library is tested for compatibility against both a Pinba server and a Pinba2 one.

Features not (yet) supported:

• Timers data misses ru_utime and ru_stime members. This is true also for timers added to PinbaClient instances

Known issues - which cannot / won't be fixed:

• lack in precision in time reporting: the time reported for page execution will be much shorter with any php code than it can be measured with a php extension. We suggest thus not to take the time reported by this package as an absolute value, but rather use it to check macro-issues, such as a page taking 10 seconds to run, or 10 times as much as another page. In the demo file doc/sample.php we showcase how to make time measurement as precise as possible
• impact on system performances: the cpu time and ram used by this implementation (which runs on every page of your site!) are also bigger than the resources used by the php extension. It is up to you to decide if the extra load added to your server by using this package is worth it or not, esp. for heavily loaded production servers
• the warnings raised when incorrect data is passed to the pinba php functions are of severity E_USER_WARNING instead of E_WARNING
• in the data reported by pinba_get_info and reported to the Pinba server, doc_size has always a value of 0. This can be worked around by using an instance of PinbaClient and calling setDocumentSize - see doc/measure_body_size.php for an example
• in the data reported to the Pinba server, memory_footprint has always a fixed value of 0 or is not reported at all. Again, using a PinbaClient instance can fix that - but there is no php function available that I know of which can report the equivalent usage of the mallinfo C call done by the php extension
• ini setting pinba.resolve_interval is not supported and most likely never will
• the default value reported to the Pina engine for the schema field is an empty string, rather than not being set at all. This results in the database table storing a value of '' instead of NULL. At the same time, sending a value of NULL makes the server-side engine re-use the last non-null value from a previous pinba packet, which seems a faulty behaviour
• the pinba_reset call does delete all exiting timers, unlike what the same function from the php extension does. Again, the upstream behaviour does feel faulty

Performances

These results are indicative of the time and memory overhead of executing 1000 function calls in a loop, and instrumenting each execution with a separate timer.

As you can see, the execution delay introduced is very small, less than 1 millisecond. The memory overhead is proportional to the number of timers added and the tags attached to each timer.

No timing:       0.00001 secs,       0 bytes used
Pinba-extension: 0.00072 secs,  280.640 bytes used
PHP-Pinba:       0.00062 secs,  412.920 bytes used

NB: weirdly enough, the php extension seems to be slightly slower on average than the pure-php implementations. Having taken a cursory look at the C code of the extension, I suspect this is because it executes too many gettimeofday calls...

In case you want to keep your code instrumented with lots of pinba_timer_start calls, and reduce the overhead of using the extension as much as possible (while of course not measuring anything anymore), you can set pinba.inhibited=1 in php.ini.

With that set, this is the overhead you can expect for "timing" 1000 executions of a function call:

No timing:      0.00001 secs,       0 bytes used
PHPPinba timed: 0.00009 secs,       0 bytes used

(tests executed with php 7.4 in an Ubuntu Focal container, running within an Ubuntu Jammy VM with 4 vCPU allocated)

Notes

Includes code from the Protobuf for PHP lib by Iván -DrSlump- Montes: https://github.com/drslump/Protobuf-PHP

Other known packages exist implementing the same idea, such as: https://github.com/vearutop/pinba-pure-php

FAQ

• Q: Can I run the polyfill in conjunction with the pinba php extension? A: yes, although I fail to see the reason why you would do that. When doing so, unless taking care to selectively disable either the php extension or this bundle (f.e. via calls to Pinba::ini_set in your code), you will get double data reported to the Pinba server

Running tests

The recommended way to run the library's test suite is via the provided Docker containers and the corresponding Docker Compose configuration. A handy shell script is available that simplifies usage of Docker and Docker Compose.

The full sequence of operations is:

./tests/ci/vm.sh build
./tests/ci/vm.sh start
./tests/ci/vm.sh runtests
./tests/ci/vm.sh stop

# and, once you have finished all testing related work:
./tests/ci/vm.sh cleanup

By default, tests are run using php 7.4 in a Container based on Ubuntu 20 Focal. The data is sent to a container running the original Pinba server - it can also be configured to be sent to a container running the Pinba2 server instead.

You can change the version of PHP and Ubuntu in use by setting the environment variables PHP_VERSION and UBUNTU_VERSION before building the containers.

You can switch the target Container used for the testsuite to the one running Pinba2 by setting the environment variables PINBA_DB_SERVER=pinba2, PINBA_SERVER=pinba2 and PINBA_PORT=3002 before starting the containers.

Testing tips

• to debug the communication between php and the Pinba server, it is possible to use tcpdump within the php container:

    sudo tcpdump udp -w packets.cap
  

  once the capture file is saved, it can be analyzed by fe. opening it in Wireshark. To improve the ability of Wireshark of decoding the protobuf-formatted messages, set up its configuration as described at https://wiki.wireshark.org/Protobuf.md. The .proto file describing the messages used by Pinba can be found at https://github.com/badoo/pinba2/blob/master/proto/pinba.proto

  As an alternative which does not require Wireshark, it is also possible to save to a file the string resulting from pinba_get_data() calls, then decode it using the protoc tool - which is included by default in the test container:

  php myTestFile > test.rawmsg
  protoc --decode=Pinba.Request tests/pinba.proto < test.rawmsg
  

License

Use of this software is subject to the terms in the license file