This virtual meetup covered several topics related to RTBkit:
1. The developer spotlight featured Nicolas Emiliani, the RTB dev team lead at Motrixi, discussing getting an RTBKit installation running.
2. Attendees learned about Motrixi's traffic which includes up to 40k queries per second from US and Canada connected to several exchanges.
3. The meetup discussed isolating the RTBKit stack using a reverse proxy, important kernel parameters, and transitioning to HTTP interfaces for RTBkit 2.0.
1 of 31
Downloaded 27 times
More Related Content
RTBkit Meetup - Developer Spotlight, Behind the Scenes of RTBkit and Intro to the Stack
2. Agenda
Developer Spotlight Nicolas Emiliani RTB Dev Lead
RTBkit Behind the Scenes
More than a Bidder The RTBkit Stack
Open Q&A
Copyright 息 2015 Datacratic Inc. All rights reserved
3. Getting an RTBKit installation running.
More info here : http://datacratic.com/site/blog/tales-rtbkit-adventure
Nicolas Emiliani, RTB dev. team Tech Leader at Motrixi.
4. A little about our traffic :
We are a mobile centered DSP
We are connected to Nexage, Rubicon, Mopub and more recently to ADX
We take up to 40k qps in peak hours from US and Canada
5. Life on the Amazon (not the rain forest)
Create a VPC (or you will regret it, I promise)
Create all your production nodes inside that VPC
Be smart about security : do not expose your nodes
unnecessarily to the internet.
Use a VPN software (ie: OpenVPN) to access the
nodes safely inside your VPC.
6. Isolating the RTBKit stack (BR loop)
Chose an HTTP reverse proxy (we use NGINX)
Set up the proxy as a gateway to your RTBKit installation, we use elastic
IP s on the NGINX nodes and allow incoming traffic only for the used
HTTP ports. Elastic IPS are cool and they will let you do all sort of handy
things.
You can use the proxy as a load balancer if you have more than one
node
You can use the proxy to automatically reply with 204s to the exchanges
in case you need to work on your installation
You can use the proxy to clone the traffic and send it to a staging or
development environment (we recently cracked how to do this with
NGINX)
8. Isolating the RTBKit PAL
As with the BR loop use a reverse proxy with an Elastic IP
9. Type of nodes
Different responsibilities Different nodes Easier scaling
10. Kernel parameters
fs.file-max=300000 : RTBKit likes to chew up your file descriptors.
net.ipv4.tcp_tw_reuse=1 : allow reusing sockets in TIME_WAIT state for
new connections.
net.ipv4.tcp_fin_timeout=15 Close connections in the TCP FIN timeout
state in 15 seconds ( default 60 )
Tweaking this parameters will prevent crashes due to exceptions related to
too many open file descriptors.
11. Use the APIs, Luke
Banker API : comes in handy when implementing pacing strategies thus
keeping logic simple in agents.
Agent Configuration Service API : is useful to implement agent
configuration related APIs as the Agent Gateway
(https://github.com/nemiliani/agent_gateway).
12. Things that will help you out
All our processes run under a watchdog, currently we are using monit
Zabbix is your friend, particularly when things go wrong.
Don't ever forget about Graphite. RTBKit is a huge black box otherwise.
Always keep an eye on your metrics. Yous should be doing it right now!,
no, really!
16. Whats taking so long?
Ancient Node.js version in platform-deps
o Upgrading requires fixing the bindings
o Lots of work, very little pay-off
Copyright 息 2015 Datacratic Inc. All rights reserved
17. Alternative?
Transition to HTTP interfaces
o Support for multiple languages
o Cleaner well defined OpenRTB protocol
o Start of RTBkit 2.0 implementation
Copyright 息 2015 Datacratic Inc. All rights reserved
18. How will it work?
Abstracted agent communication
o HTTP Bidder Interface class
o Agent config specifies bidder URL
Copyright 息 2015 Datacratic Inc. All rights reserved
19. Details
Added two new HTTP bidder examples:
o Python bidder based on tornado
o Node.js bidder
Detailed walkthrough on the wiki
Copyright 息 2015 Datacratic Inc. All rights reserved
20. Backwards compatibility?
Node.js support will be dropped for 14.04
ZMQ C++ interface will remain for now.
Will require an upgrade.
Should be fairly painless.
Copyright 息 2015 Datacratic Inc. All rights reserved
21. Platform-deps
Fork of platform-deps for RTBkit
o Upgrade dependencies to 14.04
o Cleanup of pointless dependencies
Copyright 息 2015 Datacratic Inc. All rights reserved
22. How long?
Platform-deps fork is almost completed
o Should be released in a week or two
o Work done by Michael
Currently re-writing our agents in golang
o Will iron-out kinks in the bidder interface
o Should be done in a month or two
23. More Than a Bidder
A Brief Introduction to an RTB Stack
24. When you first think about RTB ...
You think, "I need a bidder." Copyright 息 2015 Datacratic Inc. All rights reserved
25. But there is more to it than that ...
Campaign
Management
UI
Campaign
Reporting UI
ETL / Log
Processing
Offline
Optimization /
Bid Strategy
Logs / Log
Management
Ad Server
Online
Audience
Data Store
Offline Data
Prep
Monitoring Alerting 24/7 Ops
Conversion
Collection
Data
Collection /
DMP
Copyright 息 2015 Datacratic Inc. All rights reserved
26. You need an RTB Solution
Requirement Provided by RTBkit?
Reliable, high-performance core bidder system Yes
Online Impression / Click / Conversion matching Yes
Pacing No
Logging Customer has to set up and manage logs
Monitoring and Alerting Some support for Carbon and Graphite
24-7 operations and support No
Capacity planning and system evolution No
Campaign management UI integrated with the bidder No
Seats on exchanges / integration with exchanges No
ETL for campaign reporting and bidding agents No
27. Other Issues
Keep it Simple
There is an abundance of inventory. Your goal is to process enough bids to
achieve your goals. Over-engineering is expensive.
Unreliability is probably OK
o Campaigns run for a long time and buy millions of impressions. Short outages
probably don't matter
o Most partners fine with noticeable error rates (~2%)
Copyright 息 2015 Datacratic Inc. All rights reserved
28. Other Issues
Pacing and Strategy / Optimization
Hard to do in real time, at high volume of transactions, in a distributed
system
Hard to understand what data and how much data to use
Risky
o blown daily caps
o uneven or poor performance
Copyright 息 2015 Datacratic Inc. All rights reserved
29. Other Issues Ad Ops
Existing Ad Ops
Used to feature set of DSP UI
Comfortable delivering campaigns with current tool set
May be resistant to using a new system
May be resistant to trusting a new system
You may spend a lot of time on UI features and reports, not on your real-
time platform
Copyright 息 2015 Datacratic Inc. All rights reserved
30. Host an RTBkit Meetup!
Copyright 息 2015 Datacratic Inc. All rights reserved