There are many modern techniques for identifying anomalies in datasets. There are fewer that work as online algorithms suitable for application to real-time streaming data. What’s worse? Most of these methodologies require a deep understanding of the data itself. In this talk, we tour what the options are for identifying anomalies in real-time data and discuss how much we really need to know before hand to guess at the ever-useful question: is this normal?
2. Who am I?
I’m Theo (@postwait on Twitter)
I write a lot of code
50+ open source projects
several commercial code bases
I wrote “Scalable Internet Architectures”
I sit on the ACM Queue and Professions boards.
I spend all day looking at telemetry data at Circonus
3. What is real-time?
Hard real-time systems are those where the outputs of
a system based on specific inputs are considered
incorrect if the latency of their delivery is above a
specified amount.
Soft real-time systems are similar,
but “less useful” instead of “incorrect.”
I don’t design life support systems, avionics
or other systems where lives are at stake,
so it’s a soft real-time life for me.
4. A survey of big data sytems.
Traditional:
Oracle, Postgres, MySQL, Teradata,
Vertica, Netezza, Greenplum, Tableau, K
The shiny:
Hadoop, Hive, HBase, Pig, Cassandra
The real-time:
SQLstream, S4, Flumebase, Truviso, Esper, Storm
5. Big data the old way
Relational databases, both column store and not.
Just work.
Likely store more data than your “big data.”
6. Big data the distributed way
distributed systems allow much larger data sets, but
markedly change the data analytics methods
hard for existing quants to roll up their sleeves
highly scalable and accommodate growth
7. Big data the real-time way
what we do needs a different approach
the old (and even the distributed)
do not design for soft real-time complex
observation of data.
Notable exceptions are S4 and Storm.
8. So, what’s your problem?
We have telemetry...
over 10 trillion data points on near-line storage
growing super-linearly
9. Data, what kind?
Most data is numeric:
counts, averages, derivatives, stddevs, etc.
Some data is:
text changes (ssh fingerprints, production launches)
histograms
highly dimensional event streams.
10. Data rates.
Quantity of data isn’t such a big deal
okay, yes it is, but we’ll get to that later.
The rate of new data arrival makes the problem hard.
low end: 15k datum / second
high end: 300k datum / second
growing rapidly
11. What we use.
We use Esper
Esper is very powerful,
elegantly coded and
performance focused
http://www.flickr.com/photos/mcertou/
Like any good tool
that allows users to
write queries...
12. What we do with Esper
Detect absence in streams:
select b from pattern
[every a=Event -> (timer:interval(30 sec) and
not b=Event(id=a.id, metric=a.metric)]
Detect ad-hoc threshold violation:
select * from Event(id=”host1”, metric=”disk1”)
where value > 95
etc. etc. etc. [1]
13. Making the problem harder.
So, it just wasn’t enough.
We want to do long term trending
and apply that information to anomaly detection
Think: Holt-Winters (or multivariate regressions)
Look at historic data
Use that to predict the immediate future
with some quantifiable confidence.
14. How we do it.
We implemented the Snowth for storage of data. [2]
We implemented a C/lua distributed system to analyze
4 weeks of data (~8k statistical aggregates)
yielding a prediction with confidences
(triple exponential smoothing) [3]
To keep the system real-time,
we need to ensure that queries return in
less than 2ms (our goal is 100µs).
15. Cheating is winning.
Our predictions work on 5 minute windows.
4 weeks of data is 8064 windows.
Given Pred(T-8063 .. T0) -> (P1, C1)
Given Pred(T-8062 .. T0, P1) -> ~(P2, C2)
16. Tolerably inaccurate.
When V arrives,
we determine the prediction window WN we need.
If WN isn’t in cache, we assume V is within tolerances.
If WN+1 isn’t in cache,
we query the Snowth for WN, WN+1
placing in cache
Cache accesses are local and always < 100µs.
17. I see challenges
How do I
take offline data analytics techniques and
apply them online to high-volume, low-latency
event streams
quickly?
without deep expertise?
18. Thank you.
Circonus is hiring:
software engineers,
quants, and
visualization engineers.
[1] http://esper.codehaus.org/tutorials/solution_patterns/solution_patterns.html
[2] http://omniti.com/surge/2011/speakers/theo-schlossnagle
[3] http://labs.omniti.com/people/jesus/papers/holtwinters.pdf