Measuring and Logging Everything in Real Time

Transcript

1. Measuring and Logging Everything in Real-Time @BastianHofmann As the title says in the next hour we are going to talk about two very important aspects of any web application that are often overlooked in the beginning:
2. Logging 1st is logging, so that you actually now, what's happening in your application if you have to track down an error
3. Measuring And 2nd is measuring and monitoring constantly how your application is behaving
4. Many roads Of course as with many things there are many roads to accomplish that and many tools available, that help you with it. i'll not going to show you all of them, but just what we ended up using at ResearchGate. They work great for us, but depending on your use cases other tools may be more suited to your needs. The important thing to get out of this talk: take care of it
5. A few words about me before that ...
6. i work at researchgate, the social network for scientists and researchers
7. ResearchGate gives science back to the people who make it happen. We help researchers build reputation and accelerate scientific progress. On their terms. ‟ the goal is to give...
8. over 2.9 million users
9. here some impressions of the page
11. http://gigaom.com/2013/06/05/heres-how-bill-gatess- researchgate-investment-might-change-the-world-for-the- better http://venturevillage.eu/researchgate
12. have this, and also work on some cool stuff
14. in addition to this i also speak frequently on conferences throughout the world
15. work and live in berlin
16. Questions? Ask by the way, if you have any questions throughout this talk, if you don't understand something, just raise your hand and ask. probably my fault anyways since i spoke to quickly or my accent was too bad
17. Logging So 1st: Logging
18. For when something goes wrong
19. server error log access log debug logs slow query log ... Let's start with a simple setup: we have one server with an apache and some database on it. Of course all of these services are writing logs somewhere
20. $ ssh my_simple_server getting at these is very simple, just ssh into the server
21. $ tail -f error.log $ cat error.log | grep Exception and use your favorite unix command line tools to view the logs
22. Error Logs Let's look at the application related logs a bit more closely, for example the ... as it says, Apache and PHP log all errors that happen in your application there
23. ErrorLog /var/logs/apache/error.log The location is specified in your Apache configuration somewhere like this
24. and the user can get then such a nice error page, looking at this screenshot: first you may notice is the http response code displayed there
25. HTTP Response Codes http://www.w3.org/Protocols/rfc2616/rfc2616- sec6.html Please read up on it and choose correct ones for your application. helps you greatly with filtering the important stuff from the unimportant later on: a 503 may be much more serious then a 404
26. on the screenshot here you also see some additional codes and ids displayed that help to identify the error in your system later on
27. Log additional info of course you should not only display this information but also write it to your log files, you can easily do this in your custom error handler, and if you need to find the error that resulted in this error page you can just grep for the error code
28. Not only error codes
29. Request Information
30. Access Logs Another important log is the access log, which logs each incoming request
31. 192.168.56.1 - - [09/Jul/2012:19:18:19 +0200] "GET /rg_trunk/webroot/c/af10c/ images/template/rg_logo_default.png HTTP/ 1.1" 200 882 "http://devm/rg_trunk/webroot/ directory/publications/" this is what it usually looks like
32. LogFormat "%h %l %u %t \"%r\" %>s %b" custom CustomLog /var/logs/apache/access.log custom http://httpd.apache.org/docs/2.2/mod/ mod_log_config.html#logformat and you configure it in apache like this, there is already lots of information from the request you can put in there, like url, response code, request time etc
33. LogFormat "...\"%{referer}i\" \"%{user- agent}i\" %{session_id}o %{account_id}o..." custom but back to apache: these notes then can be referenced in your log format like this
34. Debug Logs next log type that your application can write are plain debug logs that are not related to errors
35. Fingers Crossed because logging everything like sql queries etc can result in huge amount of log files, and in 99% of requests you are not interested in them you should have a look at fingers crossed handlers for logging: you just put your logs in this handler but it does not log them directly to a file, only when a certain condition was met, like a threshold all already written logs and all further logs are written to a file
36. Log in a structured way one thing that can help you greatly with managing huge amount of logs with lot's of different additional information in it is logging in a structured way (not only for debug but also for error logs)
37. JSON http://www.ietf.org/rfc/rfc4627.txt in my opinion a good format is json, since it is still human readable
38. Logs from other services but of course your application probably is a bit more complex, soa anyone? so you have also other services somewhere logging something
39. web server http service http service http service http service user request log log log log log the setup may look like this, a request comes to a webserver and your php application on there calls other services. each of them have their own logs. for example the http service there. now if an error happens in the http service we are probably going to display an error as well in our php app. but how can we identify then which exception on the http service lead to the displayed error on the web server?
40. Correlation / Tracing ID a nice way of doing this in a generlized, non-custom way is by using a common correlation or tracing id
41. web server http service http service http service http service create unique trace_id for request user request trace_id trace_id trace_id trace_id log log log log log so when the user request first hits your system you generate a unique trace_id for this request and you then pass this to all underlying services through an http header.
42. X-Trace-Id: bbr8ehb984tbab894
43. web server http service http service http service http service create unique trace_id for request user request trace_id trace_id trace_id trace_id log log log log log everyone then puts these tracing id in every logs they write. so if you have a tracing id you can easily find all logs for this request by just greping for the trace_id
44. but... usually your application does not look like this
45. but more like this: you have multiple servers lying around
46. which means ssh-ing into all of these servers (dsh...) and then grepping for information over multiple logs (access log, error log, debug logs, ...) can become quite tedious and time consuming
47. Aggregate the logs in a central place so to make it easier: work on a central log management and ...
48. Make them easily full-text searchable also ...
49. Make them aggregate-able and make different kind of logs aggregateable
50. Always Log to file but whatever you do and whatever we are going to talk about...
51. Seriously...
52. Always Log to file because this is your backup when your central log management solution will fail (network errors, etc). and i'll guarantee you, it will fail sometime, probably at the worst moment
53. Directly to a database first naive approach to central log management is to log directly from you application to a database (you saw the mongoDbHandler from Monolog)
54. webserver webserver webserver DB setup would look like this: everything in one place, easily searchable, great, finished
55. Disadvantages not quite, it has some disadvantages
56. Database is down? what happens if
57. Database is slow?
58. Database is full?
59. How to integrate access logs?
60. Influences application performance because of all the database queries you are doing, and chances are if there are problems on your platform and multiple errors are occurring, directly writing to a database will make your problems worse
61. Frontend? also there is still no frontend to easily search and monitor exceptions, of course there are sql clients, phpmyadmin, rockmongo etc, but they are multi purpose tools and not really made for displaying and filtering logs
62. Better solutions? are there ...
63. graylog2 http://graylog2.org/ on tool that I quite like for this is called graylog2
64. Full text search it comes out of the box with a very nice interface to do...
65. Structured Messages in ...
66. Metrics & Alarms and offers metrics, alarms on your logs
67. that's what the interface looks like, you'll see this in action in a minute
68. https://github.com/bashofmann/ vm_graylog2_amqp_logstash_apache
69. Graylog2 UDP GELF Messages elasticsearch webserver webserver webserver the easiest approach is to just send the messages in a special format (gelf, supported by a monolog handler) with a upd request from your app servers to graylog which stores them in elasticsearch. udp is quite nice since it is fire and forget, so it does not influence our application too much (if you reuse connections and don't create a new one for each log)
70. { "version": "1.0", "host": "www1", "short_message": "Short message", "full_message": "Backtrace here\n \nmore stuff", "timestamp": 1291899928.412, "level": 1, "facility": "payment-backend", "file": "/var/www/somefile.rb", "line": 356, "_user_id": 42, "_something_else": "foo" } a typical graylog gelf message looks like this, some default fields, and user specific fields prepended with an underscore
71. Disadvantages but ...
72. Graylog/ elasticsearch is down? still if graylog/es is down you are using logs
73. Graylog/ elasticsearch is full?
74. Packet loss udp has the disadvantage that because of network errors your logs possibly don't arrive
75. Graylog2 elasticsearch webserver webserver webserver AMQP GELF GELF GELF GELF better approach, put a queue in between, also good for load balancing
76. Don't influence your application by logging but still if you push something in a queue you are still influencing your production system unnecessarily
77. Always Log to file and remember you should log to file anyways
78. logstash http://logstash.net/ enter the next tool we are using, logstash is a very powerful tool to handle log processing
79. input filter output basic workflow is that you have some input where logstash gets log messages in, on this input you can execute multiple filters that modify the message and then you can output the filtered message somewhere
80. Very rich plugin system to do this it offers a very large and rich plugin system for all kinds of inputs, filters and outputs, and you can also write your own
81. Graylog2 elasticsearch webserver webserver webserver AMQP log log log logstash logstash logstash so our setup would look like this, logstash instances on each app server take the local log files, parse and filter them and send them to the queue
82. Measuring now that you have all the log messages in a central, searchable, aggregateable place: lets go to measuring
83. For knowing what happens so that you know what actually happens in your application and in your system
84. Technical Metrics and with that i mean technical metrics like load on your servers, available memory, requests, exceptions, response times (a bit more on that later)
85. Business Metrics but also and probably even more important, business metrics
86. Define your KPIs or so called key performance indicators, example: signups, profile views, messages, uploaded images, average volume per sale, ... it really depends on your application
87. Graphite http://graphite.wikidot.com/ again there are many tools available to collect and display these metrics: one i want to highlight is graphite
88. webserver webserver webserver graphite UDP easiest way to use it is just sending udp requests with the metric you want to count to graphite, but if you're collection and measuring lots of stuff this could put quite a high load on graphite
89. StatsD https://github.com/etsy/statsd/ etsy wrote a great small tool called statsd to help you with that
90. webserver webserver webserver graphite UDP statsd you can just hook it in between your servers and graphite as a small load balancing daemon. it receives all messages, aggregates them and then flushes them in a certain interval, like very 10 seconds to graphite
91. webserver webserver webserver statsd statsd statsd graphite aggregated UPD message of course you also can run one statsd on every webserver, so that your statsd gets not overloaded
92. webserver webserver webserver statsd statsd statsd graphite aggregated UPD message statsd and you even can arrange different statsd in a row
93. Metric? what is a metric: three different types
94. Counters counters (how often did something happen)
95. Timers timer (how long did something take)
96. Gauges gauges (collect arbitrary values (amount of a sale)
97. graphite comes with a powerful interface where you can plot and aggregate this data into graphs and perform different mathematical functions on it to get it exactly the way you want to display your data
98. https://metrics.librato.com/ and if you don't want to be bothered with setting graphite up, you can also use hosted solutions like librato (statsd comes with different backends to push data to)
99. http://mashable.com/2010/10/13/etsy-office-pics/ and if you have all of this, you should make your graphs really visible, here a picture from the etsy office
100. Remember: Response Times last but not least:
101. Easiest approach: Get it out of the access log
102. 192.168.56.1 - - [09/Jul/2012:19:18:19 +0200] "GET /rg_trunk/webroot/c/af10c/ images/template/rg_logo_default.png HTTP/ 1.1" 200 882 "http://devm/rg_trunk/webroot/ directory/publications/" this is what it usually looks like
103. Is this actually what we want to measure? Is this really the response time the user sees?
104. External Resources CSS JavaScript Images Slow connections Slow Browsers SSL Handshakes DNS resolving so for the users the actual response time is actually when all this finished loading
105. That is what we want to measure!
106. https://github.com/lognormal/boomerang
107. browser JS: boomerang logstash trackingServer access log requests tracking image with timing information as query parameters graphite statsd boomerang collects from various sources load times of the current page (e.g. window.performance api), and with them requests a tracking image from a tracking server you specify (should be on another domain, so that you don't send unnecessary cookies there). on this tracking server a logstash can parse the access logs and then send the collected timers via statsd to graphite
108. http://twitter.com/BastianHofmann http://profiles.google.com/bashofmann http://lanyrd.com/people/BastianHofmann http://speakerdeck.com/u/bastianhofmann https://github.com/bashofmann https://www.researchgate.net/profile/Bastian_Hofmann/ mail@bastianhofmann.de thanks, you can contact me on any of these platforms or via mail