Design Review Best Practices - SREcon 2014


Published on

Design reviews are the foundation for a successful product or feature launch. In this session we will broach a few of the critical questions an SRE asks during the design review process to ensure the design and deployment will result in a sustainable system. We will cover real world examples of the pitfalls of not engaging the operations/infrastructure team early in the process.

Published in: Internet
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide
  • There are a lot of tools that promise to help with ETL. Many of them produce some kind of job script or war file or similar that has no logging, or predictable error reporting, or clean stop/start. They may dissolve into chaos if the data is a little bit wonky. Complex pipelines of data transformations are very hard to test, replicate, and debug.
  • Red buttoning is one of the easiest ways to control your application when an upstream service is misbehaving. The button is there to completely bypass any code reliant on the service, and continue the application flow.

    These buttons should be easy to manipulate, maybe in a CMS, or in a configuration file that is periodically re-read by the application. The intent is to not have to restart your service because someone else is having an outage.

    Red button should be a manual process; automatic graceful degradation with hold down will take care of transient errors.

    A semi-permanent use of a red button is when the upstream provider has changed their API, and your application has not been prepared; red button the feature until the app is fixed and updated.
  • The last thing you want to do with a running application in distress is dig around hoping you have the software used to build the farm originally. Keep application platforms updated; don’t let any environments get behind.
  • Design Review Best Practices - SREcon 2014

    1. 1. Design Review Best Practices Mandi Walls Technical Practice Manager, Chef SREcon 2014
    2. 2. whoami • Mandi Walls • Traveling Sysadmin for Chef • @lnxchk
    3. 3. Why Design Reviews? • Your opportunity to get a look at an incoming project • Hopefully early enough to have influence! Talk to development before it’s too late to make changes to support Ops
    4. 4. What You Want • Maximize the conversation • Get a good feel for important details • Don’t get too in the weeds, focus on what alleviates the most headaches
    5. 5. Set Your Goals • Initial deployment of the application • Runtime management and day-to-day operations • Releases and upgrades • Handling Outages
    6. 6. Have A Checklist • Combination of global requirements and needs specific to the project • Doesn’t have to be exhaustive, or hit everything the first time • Sample: • Sets goals by layer • You can also set goals by activity • It’s a living document • Make changes as you learn, launch more projects
    7. 7. Some Application FE Web Tier LB App Tier External Services BIData Tier
    8. 8. Investigating Layer by Layer • Different components will require different focus! • Some topics will hit all layers
    9. 9. Frontend Entry Point • To look at during deploy • Gathering all FQDNs • Determining SSL Requirements • Looking at Geographic Topologies
    10. 10. FE Gotchas: Rewrite Rules, Divided Farms • Sophisticated load balancing equipment allow for lots of cleverness • Documentation of intent is key • Later migrations, requirements changes are more difficult /foo /bar
    11. 11. Web Tier • Server type and version • Access methods to other tiers • Caching? • Locations • Modules and their configurations
    12. 12. Web Insanity: Custom HTTP Servers
    13. 13. Application Tier • Basics: server type and version, port requirements, protocols • Outbound connections to outside services • User affinity • Necessary libraries and their release cycles
    14. 14. App Tier Gotchas • Things that affect scaling, replacement of hosts • Whitelisting! Why!? • What changes require restarts? • New FE connections? New BE connections to cache? To DB?
    15. 15. Data Tier – Our Data • In multi-location topologies, where is the master? • Is there whitelisting or other app authorization? • What are compliance/reporting requirements? • SOX? HIPAA? PCI? • Will schemas be versioned?
    16. 16. Doing Horrible Things to Data • Premature optimization can hamstring a project • Are the data tiers susceptible to queuing? • Management of data should be automated
    17. 17. Data Tier – Imports and ETL • Schedules • Measuring successful load • Contact information for upstream provider • Storage locations and intended size
    18. 18. Operationalizing ETL • Some of the fiddliest fiddly bits
    19. 19. Data Tier – Exports, Reporting, BI • Packaging • Scheduling • Live replica versus full dump and export • Should have no impact on production operation of the datastore!
    20. 20. External APIs - Consumer • Relying on services you don’t own is tricky business • Account name and owner • SLA of service provider • Contact info for outages
    21. 21. Graceful Degradation • What does the App do if the external service is unavailable? • Shouldn’t hang • Should log some message • Should have some reasonable timeout, plus a hold down • Users shouldn’t notice
    22. 22. Red Button • Can you completely disconnect the code from a broken service? • Who can make the call?
    23. 23. Gotchas for All Tiers • Host firewalls • Application user accounts – in the central service? • Logging policy • Metrics collection • Backups • OS Versions
    24. 24. Alerts • A stacktrace is not an alertable event • Log hygiene is important – what’s in the prod logs needs to be useful • If it’s worth writing to disk, someone should know what to do with it 2014-06-01:16:44:00:UTC ERROR: $$$$$$$$ 2014-06-01:16:44:07:UTC ERROR: Host Not Found
    25. 25. Releases • Schedule – Over night? Weekends? Rolling continuous deploy? • Getting operations requirements into Dev cycle • Security requirements • Upgrades and patches • Service restarts • Graceful horizontal restarts • Full-zone downtime
    26. 26. Nightmare: Attaching Prod App to Dev DB • All artifacts should ship production ready • Localized configurations ready in advance • Lean on config management tools for other environments
    27. 27. Software Library • Set organizational standards • Rely on your OS vendors when possible • Pre-check that all dependencies are available in all repos for all environments • Suppress the urge to build special packages for absolutely everything • Streamlined, repeatable, reliable installs
    28. 28. Spelunking Through History • Keeping up with OS releases makes later scaling, replacement of dead hosts smoother • Always be current or no more than one release back in all envs
    29. 29. Performance and Tuning • Hard to be concise with new projects • Performance regression should be part of the testing process! • Know in advance what tunables exist, what their indicators are • Heap • GC • Network stacks • In-memory caching
    30. 30. Outage Planning • Known failure patterns and indicators • Who is oncall from Dev? • Who is oncall from Product or other decision-making stakeholders?
    31. 31. The Dark Art of Healthchecks • At some point, there should be a check that works back through the whole stack for status • These can be expensive • You have to know, when you hit the FE, that it is able to serve all the components of the app
    32. 32. Static Failover Pages • Mechanism for not serving up 500s or blanks during downtime • Set to non-cachable, show a picture, give a better experience
    33. 33. Other Team Info • Know who’s who • Dev teams should have some sort of oncall in case of problems Ops can’t fix • Intake process for bug reports, feedback, getting Ops issues fixed • Like those crazy log messages! • Get invited to team meetings! Go occasionally!
    34. 34. Learn From Experience • Over time, the infrastructure build out will get easier • Create standards so you don’t have to investigate all dark corners • Update your checklist periodically
    35. 35. Thanks!