SCYLLA: NoSQL at Ludicrous Speed. 主讲人 :ScyllaDB 软件工程师贺俊

Transcription

1 SCYLLA: NoSQL at Ludicrous Speed 主讲人 :ScyllaDB 软件工程师贺俊

2 Today we will cover: + Intro: Who we are, what we do, who uses it + Why we started ScyllaDB + Why should you care + How we made design decisions to achieve no-compromise performance and availability

3 Introduction + Founded by KVM hypervisor creators + Q Pivot to the database world + Q Decloak during Cassandra Summit 2015, Beta + Q General Availability + Q First Scylla Summit: 100+ Attendees + Q Completed B round + $25MM in funding + HQs: Palo Alto, CA; Herzelia, Israel employees, hiring!

4

5 Scylla benchmark by Samsung op/s

6 What we do: Scylla, towards the best NoSQL + > 1 million OPS per node + < 1ms 99% latency + Auto tuned + Scale up and out + Open source + Large community (piggyback on Cassandra) + Blends in the ecosystem- Spark, Presto, time series, search,..

7 Where Scylla is deployed?

8 Today we will cover: + Intro: Who we are, what we do, who uses it + Why we started ScyllaDB + Why should you care + How we made design decisions to achieve no-compromise performance and availability

9 Why we started Scylla? + Originally it was about performance/efficiency only + Over time, we understood we can deliver more: + SLA between background and foreground tasks + Work well on any given hardware {back pressure} + Deliver consistent, low 99th percentile latency + Reduction in admin effort + Low latency under the face of failures (hot cache load balancing) + High observability

10 Cassandra Scylla Throughput: Cannot utilize multi-core efficiently Scales linearly - shard-per-core Latency: High due to Java and JVM s GC Low and consistent - own cache Complexity: Intricate tuning and configuration Auto tuned, dynamic scheduling Admin: Maintenance impacts performance SLA guarantee for admin vs serving

11 Today we will cover: + Intro: Who we are, what we do, who uses it + Why we started ScyllaDB + Why should you care + How we made design decisions to achieve no-compromise performance and availability

12 Case study: Document column family Outbrain is the world s largest content discovery platform. Over 557 million unique visitors from across the globe. 250 billion personalized content recommendations every month.

13 Outbrain: Cassandra plus Memcache Read Microservice Write Process First read from memcached, go to Cassandra on misses. Pain: 1) Stale data from cache 2) Complexity 3) Cold cache -> C* gets full volume

14 Scylla/Cassandra side by side deployment Read Microservice Write Process Writes are written in parallel to C* and Scylla Reads are done in parallel: 1) Memcached + Cassandra 2) Scylla (no cache at all)

15 Scylla (w/o cache) vs Cassandra + Memcached Cassandra s latency Scylla s latency Requests/ Minute AVG Latency Max Latency Scylla Cassandra Diff% 12,000, ,000 (memcache handles 11,500,000) 24X 4 ms 8 ms 2X 8 ms 35 ms 3.5X Hardware 9 machines 30+9 machines 4.3X 15

16 What does it mean for a non Cassandra user? + Throughput, latency and scale benefits + Wide range of big data integration: {Kariosdb, Spark, JanusGraph, Presto, Kafka, Elastic} + Best HA/DR in the industry. + Stop using caches in front of the database + Consolidate HBase, Redis, MySQL, Mongo and others

17 Assorted Quotes

18 Today we will cover: + Intro: Who we are, what we do, who uses it + Why we started ScyllaDB + Why should you care + How we made design decisions to achieve no-compromise performance and availability

19 Design decisions: #1 The trivials

20 Design decisions: #2 Compatibility SSTable file format Configuration file format CQL language CQL native protocol JMX management protocol Management command line

21 Double cluster - Migration w/o downtime App CQL proxy Cassandra Scylla

22 Design decisions: #3 All things async

23 Design decisions: #4 Shard per core Threads Shards

24 SCYLLA DB: Network Comparison Traditional stack Scylla sharded stack queue Cassandra queue queue queue queue threads Kernel Scheduler Kernel Memory TCP/IP Application Application Application Core Database TCP/I TCP/I TCP/I P P P queue queue queue queue queue queue Task Scheduler queue queue queue queue Task Scheduler queue queue queue queue Task Scheduler queue queue queue smp queue Task Scheduler queue queue smp queue queue smp queue smp queue DPDK DPDK DPDK Userspace Userspace Userspace Userspace NIC Queues Kernel Kernel Kernel (isn t NIC (isn t NIC (isn t involved) NIC Queue involved) NIC Queue involved) Queue Queue

25 Scylla has its own task scheduler Traditional stack Scylla s stack Thread Thread Thread Thread Thread Thread Stack Thread Stack Thread Stack Stack Stack Stack Stack Stack Thread is a function pointer Stack is a byte array from 64k to megabytes Promise Promise PromiseTask PromiseTask PromiseTask Task Task Promise Promise PromiseTask PromiseTask PromiseTask Task Task Promise Promise PromiseTask PromiseTask PromiseTask Task Task Promise is a pointer to eventually computed value Task is a pointer to a lambda function Scheduler Scheduler Scheduler Scheduler Scheduler Promise Promise PromiseTask PromiseTask PromiseTask Task Task CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU

26 SCYLLA IS DIFFERENT p Thread per core p Lock-free p Task scheduler p Reactor programing p C++14 p Multi queue p Poll mode p Userspace TCP/IP p NUMA friendly p Log structured allocator p Zero copy p DMA p Log structured merge tree p DBaware cache p Userspace I/O scheduler

27 Scylla vs C* latency by Kenshoo

28 Design Decision: #5 Unified cache Cassandra Scylla Key cache Row cache On-heap / Off-heap Complex Tuning Unified cache Linux page cache SSTables SSTables

29 Design decisions: #6 I/O scheduler Cassandra Streaming configuration

30 Scylla I/O Scheduling Max useful disk concurrency Query Queue Commitlog Queue Userspace I/O Scheduler Disk Compaction Queue No queues I/O queued in FS/device

31 I/O scheduler result by Kenshoo

32 Design Decision: #7 Workload conditioning Adjust priority Memory Monitor Commitlog Memtable WAN Adjust priority Compaction Seastar Scheduler SSD Compaction Backlog Monitor Query Repair CPU

33 Workload Conditioning in practice Disk can t keep up: workload conditioning will figure out the right request rate

34 Upcoming releases + Enterprise release, based on May 2017 Counters New intra-node sharding algorithm SStableloader from 2.2/3.x Debian Sep 2017 Materialized views Execution blocks (cpu cache optimization which boost performance) Partial row cache (for wide row streaming) Heat Weighted Load Balancing

35 Scylla Beyond Cassandra Core database Vertical Horizontal

36 Q&A Resources github.com/scylladb/scylla youtube.com/c/scylladb slideshare.net/scylladb

37 THANKS SCYLLA: NoSQL at Ludicrous Speed