observe€¦ · •substantial data volumes (e.g. 500gb-1tb/day ingest is common) •long retention...
TRANSCRIPT
ObserveFOUNDED : NOV 2017 SAN MATEO, CAFOUNDERS : ARISTA,FACEBOOK,SNOWFLAKE,SPLUNK,WAVEFRONTFUNDING : $15M SERIES A – SUTTER HILL VENTURESTEAM SIZE : 15VISION : TURN THE WORLDS MACHINE DATA INTO INFORMATIONMISSION : BECOME THE MARKET LEADER IN OBSERVABILITY
WE NOW LIVE IN A DIGITAL ECONOMY.
IT PROBLEMS ARE BUSINESS PROBLEMS.DIGITAL BUSINESSES LOSE ON AVERAGE $100,000 PER HOUR.
DIAGNOSING PROBLEMS IS A NIGHTMAREDEVOPS ENGINEER IS #2 ON THE LIST OF ‘BEST JOBS IN AMERICA’
150,000+ JOB VACANCIES AT THE END OF 2018.
TOO MANY TOOLS. TOO MANY SILOS.WITHOUT A SUPERHERO YOU CAN’T SEE THE BIG PICTURE.
ITS EXPENSIVE.EXPENSIVE TO BUY. EXPENSIVE TO MAINTAIN.
bserveO Introducing
DON’T MONITOR. OBSERVE.Mon i tor ingIs The Application Running ?
Ob serv a bil ityWhy Is The Application Running This Way ?
ITS ALL ABOUT THE CLUE.TIME TO INVESTIGATE IS THE BIGGEST VARIABLE.
DATA PLATFORM
OBSERVE – KEY COMPONENTS.
INGEST LOGS, METRICS, TRACES… AND ANYTHING ELSE!
OBSERVABILITY TOOLSUNDERSTAND & INVESTIGATE SYSTEM BEHAVIOR
DATA PLATFORM – KEY CONCEPTS.
OBSERVABILITY TOOLSUNDERSTAND & INVESTIGATE SYSTEM BEHAVIOR
OBSERVATION FIRE HOSEALL EVENTS THAT HAPPENED (LOGS, METRICS & TRACES).
EVENT STREAMSIMPORTANT EVENTS THAT HAPPENED (CPU METRICS, CONTAINER LOGS…)
THINGS YOU WANT TO ASK QUESTIONS ABOUT (SERVERS, PODS…)RESOURCES
DATA
PLAT
FORM
OBSERVABILITY TOOLSUNDERSTAND & INVESTIGATE SYSTEM BEHAVIOR
10X FASTER ‘MEAN TIME TO CLUE’.
DATA PLATFORMINGEST LOGS, METRICS, TRACES… AND ANYTHING ELSE!
OBSE
RVAB
ILITY
SEARCH ANALYTICS DASHBOARDS ALERTSTIME MACHINETOOL
S
DATA PLATFORMINGEST LOGS, METRICS, TRACES… AND ANYTHING ELSE!
OUR BUSINESS MODEL IS BASED ON ANALYZING DATA.NOT INGESTING DATA.
10X LOWER COST.
$20K
$30K
$10K
$0K
$81,000
$3000
PRIC
E PE
R YE
AR50
GB
/ DAY
10X LOWER COST.
$29,000
$400
Data Management for Observability
• Most vendors build proprietary data stores for their primary use-case:• Time-series databases (e.g. Datadog, Wavefront)• Full-text search engines (e.g. Splunk, Elastic)
• Tend to be highly optimized for specific access patterns• Reading a range of a time-series (e.g. CPU on www1 for last 24 hours)• Needle in a haystack (e.g. “NullPointerException”)
Snowflake for Observability
• Principal differentiator of Observe is navigation and correlation of previously siloed data
• Most queries are just generic relational queries!• “Count the number of NullPointerExceptions over last 24 hours,
group by affected customer,and show me the top 10”
• Starting from first principles:• Want a massively scalable, columnar, scan-based OLAP database• Corroboration: Honeycomb and Scalyr both settled on this requirement too
Build or Buy?
• We chose BUY. Focus our limited resources on customer value.
(circa Mar 2018)
Properties of Observability Use-Case
• Substantial data volumes (e.g. 500GB-1TB/day ingest is common)• Long retention required in some cases (e.g. 1yr for PCI compliance)• Assume 100-1000s TB total storage per typical customer
• Recent data accessed way more often than historical (last 1h-24h)• Machine data is a mess!!• Unstructured logs, unresolved entities, sloppy formats, etc.• Too much effort to structure up-front: “ELT”
• Troubleshooting use-case is highly ad-hoc and interactive• We target <2s time to first result for most interactions
How we leverage Snowflake today
• Snowflake is embedded, transparent to the customer• Our Snowflake account, one database per customer
• Separation of storage and compute is crucial to our business model• Cheap ingest, usage-based pricing
• Multi-tenant: Shared pool of warehouses for high utilization• Tables generally clustered by TIME• Enables effective partition pruning for recent data (e.g. last 24 hours)• Conveniently, Observability data naturally arrives clustered by time
• Modeled datasets stored in distinct tables• e.g. Logs and metrics in different tables, less data to scan
How we leverage Snowflake today
• Heavy use of VARIANT type and support for semi-structured data• Lots of machine data arrives as JSON records (e.g. AWS CloudTrail Logs)
• Joins, joins, and more joins• Fundamentally, how we correlate disparate data sources
• Window analytic functions for temporal aggregation• How we materialize “Resources” from Event data
• UDFs for log parsing, string processing
• Future use: Data sharing for data import/export
Closing Thoughts
My eBay review of Snowflake:“A+++++++++ prompt shipping, product as described, would buy again”
• Snowflake has been exceptionally versatile for building a product for the Observability use-case• Buy (vs. Build) decision has enabled us to focus on delivering an end-
to-end solution rather than reinventing the wheel at every step
Looking for Development Partners!
• If you are an Elastic/Sumo Logic/Splunk/Datadog user interested in migrating to Snowflake, come talk to us!
THANK YOU
Observe