Navigating the Large Language Model choices_Ravi Daparthi
How to squeeze AWS costs
1. How to squeeze AWS costs
Jacek Migdał
Warsaw, AWS User Group, 2015-09-24
2. Who am I? The Cloud Juggler!
Senior Software Engineer at
Tripled engineering headcount in Europe!
We do logs in cloud:
● supercharged grep | awk in the browser
● Google for your internal infrastructure
Many TB / day / client
6. Challenge #1: Know your resources
Run and forget.
Problem both in development
and production.
7. Solution #1: Know your resources
Separate dev and prod:
● Separate AWS accounts.
● Consolidated billing.
● Simple to monitor and delete resources.
● Additional benefits: isolation and
security.
8. Solution #1: Know your resources
Monitor proactively:
● Tag and name resources in production.
● Setup monitoring rules.
● Get alert if you go over the limit.
● Either AWS or 3rd party (CloudHealth).
9. Solution #1: Know your resources
Shutdown by default:
● Terminate development resources after
business hours.
● Notify by email.
● Need to mark them to persist them.
10. Challenge #2: Server capacity planning
Majority of the costs are EC2.
Some planning is needed to be
cost efficient.
11. Solution #2: Server capacity planning
Best way to run machines:
● Commit for a year, get 40% discount ->
Reserved Instance.
● Use auto-scaling and spot.
● Replace with S3, DynamoDB, ...
12. Solution #2: Server capacity planning
Determine limiting metric and pick:
● Instance family (c3, m3, r3, …).
● Right-size instance type and cluster size.
Initially guess than benchmark.
Optimize the code!
13. Challenge #3: I have 100 items on my plate.
We are too busy and cost is yet
another thing.
We don’t have time to do
everything.
14. Solution #3: I have 100 items on my plate.
Rhythm:
● Buy RI regularly (e.g. monthly).
● Inspect aws resources regularly.
● Setup cost alerts at different levels.
15. Solution #3: I have 100 items on my plate.
Standardize and simplify:
● Instance types: fewer is easier for RI.
● Divide system into smaller parts by tags.
● Single owner with fallback.
16. Solution #3: I have 100 items on my plate.
Define and measure:
● usage metrics: e.g. page views
● potential optimizations impact vs. effort
If we have data, we will make the right
decision.
If we have opinions, let's go with mine.