1 of 13

Love cloud cost savings ? Let’s talk !

rolland @ rubrik

2 of 13

For the next 30 mins or so …

Efforts of 3 engineers (+ yours truly), reduced costs by 50%, under 3 months

Some answers to:

“How many people do I need ? How much time do I spend ?”

“Do I need to re-write my stuff ?” “Will I break my SLOs ?”

“Is this even relevant for me ?”

3 of 13

4 of 13

About us …

rubrik’s products & solutions

Team mission & goals

Services & capabilities offered

Also see Proactive, Real-time Monitoring and Alerting for Customer Engagement

5 of 13

Workflow under consideration

Pipeline #1�Custom code: data prep, extractors

Pipeline #2�Generic extract & load

Yeah, these are ETL pipelines

6 of 13

Where’s the problem ?

More data for decisions + Business growth (Yes, it’s always scale !)

Where were we heading ? “Can you reduce costs by 90% ?”

“What’s the plan ?” Measure, analyze, correct (repeat)

Cost analysis tools: CloudHealth + something custom (why ?)

7 of 13

Cost observability

8 of 13

Cost attribution

9 of 13

Problems, insights & solutions

Data not used ? no ROI => stop processing !

Pipeline #1: Extractor compute => waste => pre-filter data

Pipeline #2: Input size skew => utilization => bucketize

Reduced cost-dominant resource: compute (doh !)

10 of 13

Where did we end up ?

70:30

45:55

Compute

Storage

$$$

11 of 13

What else is possible ?

Flexible processing (eg. on-demand) -vs- every upload “event”

Data reduction at source

Pipeline #1: generic “declarative” extractor

Infra changes & tuning: eg. spot nodes

12 of 13

Learnings

Start quick & dirty: no EMR metrics ? no problem

Low hanging fruit: top 5 extractors, top 2 tables

Make it a “team sport” - validate ROI

Preparation ? Yes, familiarity with the system helped

Misses ? “How do I reduce my extractor costs ?”

13 of 13

Hope that helped …