aboutsummaryrefslogtreecommitdiff
path: root/public/what-i-ve-learned-developing-ad-server.html
diff options
context:
space:
mode:
authorMitja Felicijan <m@mitjafelicijan.com>2023-07-08 23:25:41 +0200
committerMitja Felicijan <m@mitjafelicijan.com>2023-07-08 23:25:41 +0200
commitcd6644ea4ddc78597934ab0ef5ba50e3c3daa927 (patch)
tree03de331a8db6386dfd6fa75155bfbcea6b4feaf3 /public/what-i-ve-learned-developing-ad-server.html
parent84ed124529ffeee1590295b8de3a8faf51848680 (diff)
downloadmitjafelicijan.com-cd6644ea4ddc78597934ab0ef5ba50e3c3daa927.tar.gz
Moved to a simpler SSG
Diffstat (limited to 'public/what-i-ve-learned-developing-ad-server.html')
-rwxr-xr-xpublic/what-i-ve-learned-developing-ad-server.html126
1 files changed, 126 insertions, 0 deletions
diff --git a/public/what-i-ve-learned-developing-ad-server.html b/public/what-i-ve-learned-developing-ad-server.html
new file mode 100755
index 0000000..44e8d61
--- /dev/null
+++ b/public/what-i-ve-learned-developing-ad-server.html
@@ -0,0 +1,126 @@
1<!doctype html><html lang=en-us><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><link href="data:image/x-icon;base64,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" rel=icon type=image/x-icon><title>What I've learned developing ad server</title><meta name=description content="For the past year and half I have been developing native advertising server thatcontextually matches ads and displays them in different template forms onvariety of websites."><link rel=alternate type=application/rss+xml title="Mitja Felicijan's posts" href=https://mitjafelicijan.com/index.xml><link rel=alternate type=application/rss+xml title="Mitja Felicijan's notes" href=https://mitjafelicijan.com/notes.xml><style>body{padding:1rem;max-width:760px;background:#fff;font-family:times new roman,Times,serif;line-height:1.35rem}hr{margin-block-start:1.5rem}h1,h2,h3{line-height:initial}footer{margin-block-start:3rem}table{max-width:100%;border-collapse:separate;border-spacing:2px;border:1px solid #000;border-left:1px solid #999;border-top:1px solid #999}blockquote{font-style:italic}table thead{background:#eee}td,th{border:1px solid #000;padding:4px;border-right:1px solid #999;border-bottom:1px solid #999;text-align:left}pre{text-wrap:nowrap;overflow-x:auto;margin-block-start:1.5rem;margin-block-end:1.5rem;padding:.5rem 0;border-top:1px solid #000;border-bottom:1px solid #000}pre code{line-height:1.3em}pre,code,pre *,code *{font-family:monospace;font-size:initial!important}img,video,audio{max-width:100%}header{display:flex;flex-direction:row;gap:3rem}nav{display:flex;gap:.75rem}.pstatus-orange{background:gold}.pstatus-green{background:#9acd32}.pstatus-red{background:#cd5c5c}@media only screen and (max-width:600px){header{flex-direction:column;gap:1rem}a{word-wrap:break-word}}</style><header><nav class=main><a href=/>Home</a>
2<a href=https://git.mitjafelicijan.com/ target=_blank>Git</a>
3<a href=https://files.mitjafelicijan.com/ target=_blank>Files</a>
4<a href=/mitjafelicijan.pgp.pub.txt target=_blank>PGP</a>
5<a href=/curriculum-vitae.html>CV</a>
6<a href=/index.xml target=_blank>RSS</a></nav></header><main><div><h1>What I've learned developing ad server</h1><p>Apr 17, 2017<div><p>For the past year and half I have been developing native advertising server that
7contextually matches ads and displays them in different template forms on
8variety of websites. This project grew from serving thousands of ads per day to
9millions.<p>The system is made from couple of core components:<ul><li>API for serving ads,<li>Utils - cronjobs and queue management tools,<li>Dashboard UI.</ul><p>Initial release was using <a href=https://www.mongodb.com/>MongoDB</a> for full-text
10search but was later replaced by <a href=https://www.elastic.co/>Elasticsearch</a> for
11better CPU utilization and better search performance. This provided us with many
12amazing functionalities of <a href=https://www.elastic.co/>Elasticsearch</a>. You should
13check it out if you do any search related operations.<p>Because the premise of the server is to provide native ad experience, they are
14rendered on the client side via simple templating engine. This ensures that ads
15can be displayed number of different ways based on the visual style of the
16page. And this makes JavaScript client library quite complex.<p>So now that you know basic information about the product lets get into the
17lessons we learned.<h2 id=aggregate-everything>Aggregate everything</h2><p>After beta version was released everything (impressions, clicks, etc) was
18written in nanosecond resolution in the database. At that time we were using
19<a href=https://www.postgresql.org/>PostgreSQL</a> and database quickly grew way above
20200GB in disk space. And that was problematic. Statistics took disturbingly long
21time to aggregate. Also using indexes on stats table in database was no help
22after we reached 500 million datapoints.<blockquote><p>There is a marketing product information and there is real life experience.
23And the tend to be quite the opposite.</blockquote><p>This was the reason that now everything is aggregated on daily basis and this
24data is then fed to Elastic in form of daily summary. With this we achieved we
25can now track many more dimensions such as zone, channel and platform
26information. And with this information we can now adapt occurrences of ads on
27specific places more precisely.<p>We have also adapted <a href=https://redis.io/>Redis</a> as a full-time citizen in our
28stack. Because Redis also stores information on a local disk we have some sort
29of backup if server would accidentally suffer some failure.<p>All the real-time statistics for ad serving and redirecting is presented as
30counters in Redis instance and daily extracted and pushed to Elastic.<h2 id=measure-everything>Measure everything</h2><p>The thing about software is that we really don't know how well it is performing
31under load until such load is presented. When testing locally everything is fine
32but when on production things tend to fall apart.<p>As a solution for this we are measuring everything we can. Function execution
33time (by encapsulating functions with timers), server performance (cpu, memory,
34disk, etc), Nginx and <a href=https://uwsgi-docs.readthedocs.io/>uWSGI</a> performance.
35We sacrifice a bit of performance for the sake of this information. And we store
36all this information for later analysis.<p><strong>Example of function execution time</strong><pre tabindex=0 style=background-color:#fff><code><span style=display:flex><span>{
37</span></span><span style=display:flex><span> &#34;get_final_filtered_ads&#34;: {
38</span></span><span style=display:flex><span> &#34;counter&#34;: 1931250,
39</span></span><span style=display:flex><span> &#34;avg&#34;: 0.0066143431,
40</span></span><span style=display:flex><span> &#34;elapsed&#34;: 12773.9500310003
41</span></span><span style=display:flex><span> },
42</span></span><span style=display:flex><span> &#34;store_keywords_statistics&#34;: {
43</span></span><span style=display:flex><span> &#34;counter&#34;: 1931011,
44</span></span><span style=display:flex><span> &#34;avg&#34;: 0.0004605267,
45</span></span><span style=display:flex><span> &#34;elapsed&#34;: 889.2821669996
46</span></span><span style=display:flex><span> },
47</span></span><span style=display:flex><span> &#34;match_by_context&#34;: {
48</span></span><span style=display:flex><span> &#34;counter&#34;: 1931011,
49</span></span><span style=display:flex><span> &#34;avg&#34;: 0.0055960716,
50</span></span><span style=display:flex><span> &#34;elapsed&#34;: 10806.0758889999
51</span></span><span style=display:flex><span> },
52</span></span><span style=display:flex><span> &#34;match_by_high_performance&#34;: {
53</span></span><span style=display:flex><span> &#34;counter&#34;: 262,
54</span></span><span style=display:flex><span> &#34;avg&#34;: 0.0152770229,
55</span></span><span style=display:flex><span> &#34;elapsed&#34;: 4.00258
56</span></span><span style=display:flex><span> },
57</span></span><span style=display:flex><span> &#34;store_impression_stats&#34;: {
58</span></span><span style=display:flex><span> &#34;counter&#34;: 1931250,
59</span></span><span style=display:flex><span> &#34;avg&#34;: 0.0006189991,
60</span></span><span style=display:flex><span> &#34;elapsed&#34;: 1195.4419869999
61</span></span><span style=display:flex><span> }
62</span></span><span style=display:flex><span>}
63</span></span></code></pre><p>We have also started profiling with <a href=https://pymotw.com/2/profile/>cProfile</a>
64and then visualizing with <a href=http://kcachegrind.sourceforge.net/>KCachegrind</a>.
65This provides much more detailed look into code execution.<h2 id=cache-control-is-your-friend>Cache control is your friend</h2><p>Because we use Javascript library for rendering ads we rely on this script
66extensively and when in need we need to be able to change behavior of the script
67quickly.<p>In our case we can not simply replace javascript url in html code. It usually
68takes a day or two for the guys who maintain sites to change code or add
69?ver=xxx attribute. And this makes rapid deployment and testing very difficult
70and time consuming. There is a limitation of how much you can test locally.<p>We are now in the process of integrating <a href=https://www.google.com/analytics/tag-manager/>Google Tag
71Manager</a> but couple of websites
72are developed on ASP.net platform that have some problems with tag manager. With
73a solution below we are certain that we are serving latest version of the
74script.<p>And it only takes one mistake and users have the script cached and in case of
75caching it for 1 year you probably know where the problem is.<pre tabindex=0 style=background-color:#fff><code><span style=display:flex><span><span style=color:green># nginx ➜ /etc/nginx/sites-available/default
76</span></span></span><span style=display:flex><span><span style=color:green></span><span style=color:#00f>location</span> <span style=color:#a31515>/static/</span> {
77</span></span><span style=display:flex><span> <span style=color:#00f>alias</span> <span style=color:#a31515>/path-to-static-content/</span>;
78</span></span><span style=display:flex><span> <span style=color:#00f>autoindex</span> off;
79</span></span><span style=display:flex><span> <span style=color:#00f>charset</span> <span style=color:#a31515>utf-8</span>;
80</span></span><span style=display:flex><span> <span style=color:#00f>gzip</span> on;
81</span></span><span style=display:flex><span> <span style=color:#00f>gzip_types</span> <span style=color:#a31515>text/plain</span> <span style=color:#a31515>application/javascript</span> <span style=color:#a31515>application/x-javascript</span> <span style=color:#a31515>text/javascript</span> <span style=color:#a31515>text/xml</span> <span style=color:#a31515>text/css</span>;
82</span></span><span style=display:flex><span> <span style=color:#00f>location</span> ~<span style=color:#a31515>*</span> <span style=color:#a31515>\.(ico|gif|jpeg|jpg|png|woff|ttf|otf|svg|woff2|eot)</span>$ {
83</span></span><span style=display:flex><span> <span style=color:#00f>expires</span> <span style=color:#a31515>1y</span>;
84</span></span><span style=display:flex><span> <span style=color:#00f>add_header</span> <span style=color:#a31515>Pragma</span> <span style=color:#a31515>public</span>;
85</span></span><span style=display:flex><span> <span style=color:#00f>add_header</span> <span style=color:#a31515>Cache-Control</span> <span style=color:#a31515>&#34;public&#34;</span>;
86</span></span><span style=display:flex><span> }
87</span></span><span style=display:flex><span> <span style=color:#00f>location</span> ~<span style=color:#a31515>*</span> <span style=color:#a31515>\.(css|js|txt)</span>$ {
88</span></span><span style=display:flex><span> <span style=color:#00f>expires</span> <span style=color:#a31515>3600s</span>;
89</span></span><span style=display:flex><span> <span style=color:#00f>add_header</span> <span style=color:#a31515>Pragma</span> <span style=color:#a31515>public</span>;
90</span></span><span style=display:flex><span> <span style=color:#00f>add_header</span> <span style=color:#a31515>Cache-Control</span> <span style=color:#a31515>&#34;public,</span> <span style=color:#a31515>must-revalidate&#34;</span>;
91</span></span><span style=display:flex><span> }
92</span></span><span style=display:flex><span>}
93</span></span></code></pre><p>Also be careful when redirecting to url in your python code. We noticed that if
94we didn't precisely setup cache control and expire headers in response we didn't
95get the request on the server and therefore couldn't measure clicks. So when
96redirecting do as follows and there will be no problems.<pre tabindex=0 style=background-color:#fff><code><span style=display:flex><span><span style=color:green># python ➜ bottlepy web micro-framework</span>
97</span></span><span style=display:flex><span>response = bottle.HTTPResponse(status=302)
98</span></span><span style=display:flex><span>response.set_header(<span style=color:#a31515>&#34;Cache-Control&#34;</span>, <span style=color:#a31515>&#34;no-store, no-cache, must-revalidate&#34;</span>)
99</span></span><span style=display:flex><span>response.set_header(<span style=color:#a31515>&#34;Expires&#34;</span>, <span style=color:#a31515>&#34;Thu, 01 Jan 1970 00:00:00 GMT&#34;</span>)
100</span></span><span style=display:flex><span>response.set_header(<span style=color:#a31515>&#34;Location&#34;</span>, url)
101</span></span><span style=display:flex><span><span style=color:#00f>return</span> response
102</span></span></code></pre><blockquote><p>Cache control in browsers is quite aggressive and you need to be precise to
103avoid future problems. We learned that lesson the hard way.</blockquote><h2 id=learn-nginx>Learn NGINX</h2><p>When deciding on a web server we went with Nginx as a reverse proxy for our
104applications. We adapted micro-service oriented architecture early in the
105project to ensure when we scale we can easily add additional servers to our
106cluster. And Nginx was crucial to perform load balancing and static content
107delivery.<p>At first our config file was quite simple and later grew larger. After patching
108and adding new settings I sat down and learned more about the guts of Nginx.
109This proved to be very useful and we were able to squeeze much more out of our
110setup. So I advise you to take your time and read through the
111<a href=https://nginx.org/en/docs/>documentation</a>. This saved us a lot of headache.
112Googling for solutions only goes so far.<h2 id=use-redismemcached>Use Redis/Memcached</h2><p>As explained above we are using caching basically for everything. It is the
113corner stone of our services. At first we were very careful about the quantity
114of things we stored in <a href=https://redis.io/>Redis</a>. But we later found out that
115the memory footprint is very low even when storing large amount of data in it.<p>So we gradually increased our usage to caching whole HTML outputs of dashboard.
116This improved our performance in order of magnitude. And by using native TTL
117support this goes hand in hand with our needs.<p>The reason why we choose <a href=https://redis.io/>Redis</a> over
118<a href=https://memcached.org/>Memcached</a> was the nature of scalability of Redis out
119of the box. But all this can be achieved with Memcached.<h2 id=conclusion>Conclusion</h2><p>There are a lot more details that could have been written and every single topic
120in here deserves it's own post but you probably got the idea about the problems
121we faced.</div></div></main><footer><hr><div><h3>Want to comment or have something to add?</h3>You can write me an email at
122<a href=mailto:m@mitjafelicijan.com>m@mitjafelicijan.com</a> or catch up
123with me
124<a href=https://telegram.me/mitjafelicijan target=_blank>on Telegram</a>.</div><hr><p>This website does not track you. Content is made available under
125the <a href=https://creativecommons.org/licenses/by/4.0/ target=_blank rel=noreferrer>CC BY 4.0 license</a> unless specified
126otherwise. Blog feed is available as <a href=/index.xml target=_blank>RSS feed</a>.</footer><script src=https://cdn.usefathom.com/script.js data-site=XHQARKXP defer></script> \ No newline at end of file
generated by cgit v1.2.3 (git 2.46.0) at 2026-04-30 21:38:33 +0000