1 files changed, 0 insertions, 145 deletions
diff --git a/content/2017-04-17-what-i-ve-learned-developing-ad-server.md b/content/2017-04-17-what-i-ve-learned-developing-ad-server.md
deleted file mode 100644
index c186484..0000000
--- a/content/2017-04-17-what-i-ve-learned-developing-ad-server.md
+++ /dev/null
@@ -1,145 +0,0 @@
---
-layout: post
-title: What I've learned developing ad server
-description: Lessons I learned developing contextual ad server
-slug: what-i-ve-learned-developing-ad-server
-type: note
-date: 2017-04-17
---
-**Table of contents**
-1. [Aggregate everything](#aggregate-everything)
-2. [Measure everything](#measure-everything)
-3. [Cache control is your friend](#cache-control-is-your-friend)
-4. [Learn NGINX](#learn-nginx)
-5. [Use Redis/Memcached](#use-redismemcached)
-6. [Conclusion](#conclusion)
-For the past year and half I have been developing native advertising server that contextually matches ads and displays them in different template forms on variety of websites. This project grew from serving thousands of ads per day to millions.
-The system is made from couple of core components:
- API for serving ads,
- Utils - cronjobs and queue management tools,
- Dashboard UI.
-Initial release was using [MongoDB](https://www.mongodb.com/) for full-text search but was later replaced by [Elasticsearch](https://www.elastic.co/) for better CPU utilization and better search performance. This provided us with many amazing functionalities of [Elasticsearch](https://www.elastic.co/). You should check it out if you do any search related operations.
-Because the premise of the server is to provide native ad experience, they are rendered on the client side via simple templating engine. This ensures that ads can be displayed number of different ways based on the visual style of the page. And this makes Javascript client library quite complex.
-So now that you know basic information about the product lets get into the lessons we learned.
-## Aggregate everything
-After beta version was released everything (impressions, clicks, etc) was written in nanosecond resolution in the database. At that time we were using [PostgreSQL](https://www.postgresql.org/) and database quickly grew way above 200GB in disk space. And that was problematic. Statistics took disturbingly long time to aggregate. Also using indexes on stats table in database was no help after we reached 500 million datapoints.
-> There is a marketing product information and there is real life experience. And the tend to be quite the opposite.
-This was the reason that now everything is aggregated on daily basis and this data is then fed to Elastic in form of daily summary. With this we achieved we can now track many more dimensions such as zone, channel and platform information. And with this information we can now adapt occurrences of ads on specific places more precisely.
-We have also adapted [Redis](https://redis.io/) as a full-time citizen in our stack. Because Redis also stores information on a local disk we have some sort of backup if server would accidentally suffer some failure.
-All the real-time statistics for ad serving and redirecting is presented as counters in Redis instance and daily extracted and pushed to Elastic.
-## Measure everything
-The thing about software is that we really don't know how well it is performing under load until such load is presented. When testing locally everything is fine but when on production things tend to fall apart.
-As a solution for this we are measuring everything we can. Function execution time (by encapsulating functions with timers), server performance (cpu, memory, disk, etc), Nginx and [uWSGI](https://uwsgi-docs.readthedocs.io/) performance. We sacrifice a bit of performance for the sake of this information. And we store all this information for later analysis.
-**Example of function execution time**
-```json
-{
-  "get_final_filtered_ads": {
-    "counter": 1931250,
-    "avg": 0.0066143431,
-    "elapsed": 12773.9500310003
-  },
-  "store_keywords_statistics": {
-    "counter": 1931011,
-    "avg": 0.0004605267,
-    "elapsed": 889.2821669996
-  },
-  "match_by_context": {
-    "counter": 1931011,
-    "avg": 0.0055960716,
-    "elapsed": 10806.0758889999
-  },
-  "match_by_high_performance": {
-    "counter": 262,
-    "avg": 0.0152770229,
-    "elapsed": 4.00258
-  },
-  "store_impression_stats": {
-    "counter": 1931250,
-    "avg": 0.0006189991,
-    "elapsed": 1195.4419869999
-  }
-}
-```
-We have also started profiling with [cProfile](https://pymotw.com/2/profile/) and then visualizing with [KCachegrind](http://kcachegrind.sourceforge.net/). This provides much more detailed look into code execution.
-## Cache control is your friend
-Because we use Javascript library for rendering ads we rely on this script extensively and when in need we need to be able to change behavior of the script quickly.
-In our case we can not simply replace javascript url in html code. It usually takes a day or two for the guys who maintain sites to change code or add ?ver=xxx attribute. And this makes rapid deployment and testing very difficult and time consuming. There is a limitation of how much you can test locally.
-We are now in the process of integrating [Google Tag Manager](https://www.google.com/analytics/tag-manager/) but couple of websites are developed on ASP.net platform that have some problems with tag manager. With a solution below we are certain that we are serving latest version of the script.
-And it only takes one mistake and users have the script cached and in case of caching it for 1 year you probably know where the problem is.
-```nginx
-# nginx ➜ /etc/nginx/sites-available/default
-location /static/ {
-  alias /path-to-static-content/;
-  autoindex off;
-  charset utf-8;
-  gzip on;
-  gzip_types text/plain application/javascript application/x-javascript text/javascript text/xml text/css;
-  location ~* \.(ico|gif|jpeg|jpg|png|woff|ttf|otf|svg|woff2|eot)$ {
-    expires 1y;
-    add_header Pragma public;
-    add_header Cache-Control "public";
-  }
-  location ~* \.(css|js|txt)$ {
-    expires 3600s;
-    add_header Pragma public;
-    add_header Cache-Control "public, must-revalidate";
-  }
-}
-```
-Also be careful when redirecting to url in your python code. We noticed that if we didn't precisely setup cache control and expire headers in response we didn't get the request on the server and therefore couldn't measure clicks. So when redirecting do as follows and there will be no problems.
-```python
-# python ➜ bottlepy web micro-framework
-response = bottle.HTTPResponse(status=302)
-response.set_header("Cache-Control", "no-store, no-cache, must-revalidate")
-response.set_header("Expires", "Thu, 01 Jan 1970 00:00:00 GMT")
-response.set_header("Location", url)
-return response
-```
-> Cache control in browsers is quite aggressive and you need to be precise to avoid future problems. We learned that lesson the hard way.
-## Learn NGINX
-When deciding on a web server we went with Nginx as a reverse proxy for our applications. We adapted micro-service oriented architecture early in the project to ensure when we scale we can easily add additional servers to our cluster. And Nginx was crucial to perform load balancing and static content delivery.
-At first our config file was quite simple and later grew larger. After patching and adding new settings I sat down and learned more about the guts of Nginx. This proved to be very useful and we were able to squeeze much more out of our setup. So I advise you to take your time and read through the [documentation](https://nginx.org/en/docs/). This saved us a lot of headache. Googling for solutions only goes so far.
-## Use Redis/Memcached
-As explained above we are using caching basically for everything. It is the corner stone of our services. At first we were very careful about the quantity of things we stored in [Redis](https://redis.io/). But we later found out that the memory footprint is very low even when storing large amount of data in it.
-So we gradually increased our usage to caching whole HTML outputs of dashboard. This improved our performance in order of magnitude. And by using native TTL support this goes hand in hand with our needs.
-The reason why we choose [Redis](https://redis.io/) over [Memcached](https://memcached.org/) was the nature of scalability of Redis out of the box. But all this can be achieved with Memcached.
-## Conclusion
-There are a lot more details that could have been written and every single topic in here deserves it's own post but you probably got the idea about the problems we faced.

diff --git a/content/2017-04-17-what-i-ve-learned-developing-ad-server.md b/content/2017-04-17-what-i-ve-learned-developing-ad-server.md deleted file mode 100644 index c186484..0000000 --- a/content/2017-04-17-what-i-ve-learned-developing-ad-server.md +++ /dev/null
@@ -1,145 +0,0 @@
1	---
2	layout: post
3	title: What I've learned developing ad server
4	description: Lessons I learned developing contextual ad server
5	slug: what-i-ve-learned-developing-ad-server
6	type: note
7	date: 2017-04-17
8	---
9
10	Table of contents
11
12	1. [Aggregate everything](#aggregate-everything)
13	2. [Measure everything](#measure-everything)
14	3. [Cache control is your friend](#cache-control-is-your-friend)
15	4. [Learn NGINX](#learn-nginx)
16	5. [Use Redis/Memcached](#use-redismemcached)
17	6. [Conclusion](#conclusion)
18
19	For the past year and half I have been developing native advertising server that contextually matches ads and displays them in different template forms on variety of websites. This project grew from serving thousands of ads per day to millions.
20
21	The system is made from couple of core components:
22
23	- API for serving ads,
24	- Utils - cronjobs and queue management tools,
25	- Dashboard UI.
26
27	Initial release was using [MongoDB](https://www.mongodb.com/) for full-text search but was later replaced by [Elasticsearch](https://www.elastic.co/) for better CPU utilization and better search performance. This provided us with many amazing functionalities of [Elasticsearch](https://www.elastic.co/). You should check it out if you do any search related operations.
28
29	Because the premise of the server is to provide native ad experience, they are rendered on the client side via simple templating engine. This ensures that ads can be displayed number of different ways based on the visual style of the page. And this makes Javascript client library quite complex.
30
31	So now that you know basic information about the product lets get into the lessons we learned.
32
33	## Aggregate everything
34
35	After beta version was released everything (impressions, clicks, etc) was written in nanosecond resolution in the database. At that time we were using [PostgreSQL](https://www.postgresql.org/) and database quickly grew way above 200GB in disk space. And that was problematic. Statistics took disturbingly long time to aggregate. Also using indexes on stats table in database was no help after we reached 500 million datapoints.
36
37	> There is a marketing product information and there is real life experience. And the tend to be quite the opposite.
38
39	This was the reason that now everything is aggregated on daily basis and this data is then fed to Elastic in form of daily summary. With this we achieved we can now track many more dimensions such as zone, channel and platform information. And with this information we can now adapt occurrences of ads on specific places more precisely.
40
41	We have also adapted [Redis](https://redis.io/) as a full-time citizen in our stack. Because Redis also stores information on a local disk we have some sort of backup if server would accidentally suffer some failure.
42
43	All the real-time statistics for ad serving and redirecting is presented as counters in Redis instance and daily extracted and pushed to Elastic.
44
45	## Measure everything
46
47	The thing about software is that we really don't know how well it is performing under load until such load is presented. When testing locally everything is fine but when on production things tend to fall apart.
48
49	As a solution for this we are measuring everything we can. Function execution time (by encapsulating functions with timers), server performance (cpu, memory, disk, etc), Nginx and [uWSGI](https://uwsgi-docs.readthedocs.io/) performance. We sacrifice a bit of performance for the sake of this information. And we store all this information for later analysis.
50
51	Example of function execution time
52
53	```json
54	{
55	"get_final_filtered_ads": {
56	"counter": 1931250,
57	"avg": 0.0066143431,
58	"elapsed": 12773.9500310003
59	},
60	"store_keywords_statistics": {
61	"counter": 1931011,
62	"avg": 0.0004605267,
63	"elapsed": 889.2821669996
64	},
65	"match_by_context": {
66	"counter": 1931011,
67	"avg": 0.0055960716,
68	"elapsed": 10806.0758889999
69	},
70	"match_by_high_performance": {
71	"counter": 262,
72	"avg": 0.0152770229,
73	"elapsed": 4.00258
74	},
75	"store_impression_stats": {
76	"counter": 1931250,
77	"avg": 0.0006189991,
78	"elapsed": 1195.4419869999
79	}
80	}
81	```
82
83	We have also started profiling with [cProfile](https://pymotw.com/2/profile/) and then visualizing with [KCachegrind](http://kcachegrind.sourceforge.net/). This provides much more detailed look into code execution.
84
85	## Cache control is your friend
86
87	Because we use Javascript library for rendering ads we rely on this script extensively and when in need we need to be able to change behavior of the script quickly.
88
89	In our case we can not simply replace javascript url in html code. It usually takes a day or two for the guys who maintain sites to change code or add ?ver=xxx attribute. And this makes rapid deployment and testing very difficult and time consuming. There is a limitation of how much you can test locally.
90
91	We are now in the process of integrating [Google Tag Manager](https://www.google.com/analytics/tag-manager/) but couple of websites are developed on ASP.net platform that have some problems with tag manager. With a solution below we are certain that we are serving latest version of the script.
92
93	And it only takes one mistake and users have the script cached and in case of caching it for 1 year you probably know where the problem is.
94
95	```nginx
96	# nginx ➜ /etc/nginx/sites-available/default
97	location /static/ {
98	alias /path-to-static-content/;
99	autoindex off;
100	charset utf-8;
101	gzip on;
102	gzip_types text/plain application/javascript application/x-javascript text/javascript text/xml text/css;
103	location ~* \.(ico\|gif\|jpeg\|jpg\|png\|woff\|ttf\|otf\|svg\|woff2\|eot)$ {
104	expires 1y;
105	add_header Pragma public;
106	add_header Cache-Control "public";
107	}
108	location ~* \.(css\|js\|txt)$ {
109	expires 3600s;
110	add_header Pragma public;
111	add_header Cache-Control "public, must-revalidate";
112	}
113	}
114	```
115
116	Also be careful when redirecting to url in your python code. We noticed that if we didn't precisely setup cache control and expire headers in response we didn't get the request on the server and therefore couldn't measure clicks. So when redirecting do as follows and there will be no problems.
117
118	```python
119	# python ➜ bottlepy web micro-framework
120	response = bottle.HTTPResponse(status=302)
121	response.set_header("Cache-Control", "no-store, no-cache, must-revalidate")
122	response.set_header("Expires", "Thu, 01 Jan 1970 00:00:00 GMT")
123	response.set_header("Location", url)
124	return response
125	```
126
127	> Cache control in browsers is quite aggressive and you need to be precise to avoid future problems. We learned that lesson the hard way.
128
129	## Learn NGINX
130
131	When deciding on a web server we went with Nginx as a reverse proxy for our applications. We adapted micro-service oriented architecture early in the project to ensure when we scale we can easily add additional servers to our cluster. And Nginx was crucial to perform load balancing and static content delivery.
132
133	At first our config file was quite simple and later grew larger. After patching and adding new settings I sat down and learned more about the guts of Nginx. This proved to be very useful and we were able to squeeze much more out of our setup. So I advise you to take your time and read through the [documentation](https://nginx.org/en/docs/). This saved us a lot of headache. Googling for solutions only goes so far.
134
135	## Use Redis/Memcached
136
137	As explained above we are using caching basically for everything. It is the corner stone of our services. At first we were very careful about the quantity of things we stored in [Redis](https://redis.io/). But we later found out that the memory footprint is very low even when storing large amount of data in it.
138
139	So we gradually increased our usage to caching whole HTML outputs of dashboard. This improved our performance in order of magnitude. And by using native TTL support this goes hand in hand with our needs.
140
141	The reason why we choose [Redis](https://redis.io/) over [Memcached](https://memcached.org/) was the nature of scalability of Redis out of the box. But all this can be achieved with Memcached.
142
143	## Conclusion
144
145	There are a lot more details that could have been written and every single topic in here deserves it's own post but you probably got the idea about the problems we faced.