Hpts Keynote Pat Helland
HTPS Keynote
Before
- Storage has evolved
- (from Direct, SAN, Storage clusters + REST API)
- Computing Evolved
- Single proc -> Multi -> RPC -> SOA -> Microservices
- Computing Storage Use
- Direct File I/O
State
- Durable state + consistency (transactions/updates/message semantics)
- Session State - multi op transactions; sessions are hard @ microservices
- inside/outside data
- outside data is immutable
Careful Replacement
- disks can trash block during write: (old -> trash -> new)
- hardware fixes careful replacement issue
- transactions bundle + solve careful replacement
- workflows need careful transactional replacement now
Read your writes?
- Strong consistently - (R after W):
- Even as scales you can r after w
- Sometime delay if sick/dead (releader election)
- Weakly consistent (not R after W):
- No guarantee write update old replica - read old value
- R/W consistent SLA - skip sick/dead servers
- Cached Data
- KV pairs versioned, unsure if newest.
Different stores for diff uses! Immutability can fix these - only one version to get.
Stateful Sessions And Transactions
- Stateful sessions worked well for SOA: Distributed Transactions + Across Machines
Transactions, Sessions, Microservice
- Requests flow to request routing.
- Session state needed for cross-request TV
- Microservice Transaction are typically 1 store request
- No cross request transactions
Theres no substitute for interchangibility
Written on October 2, 2017