HBase: Extreme makeover
Vladimir RodionovHadoop/HBase architectFounder of BigBase.org
HBaseCon 2014Features & Internal Track
Agenda
About myself• Principal Platform Engineer @Carrier IQ, Sunnyvale, CA • Prior to Carrier IQ, I worked @ GE, EBay, Plumtree/BEA.• HBase user since 2009.• HBase hacker since 2013.• Areas of expertise include (but not limited to) Java,
HBase, Hadoop, Hive, large-scale OLAP/Analytics, and in-memory data processing.
• Founder of BigBase.org
What?
BigBase
BigBase = EM(HBase)
BigBase = EM(HBase)
EM(*) = ?
BigBase = EM(HBase)
EM(*) =
BigBase = EM(HBase)
EM(*) =
Seriously?
BigBase = EM(HBase)
EM(*) =
Seriously?for HBaseIt’s a Multi-Level Caching solution
Real Agenda• Why BigBase?• Brief history of BigBase.org project• BigBase MLC high level architecture (L1/L2/L3)• Level 1 - Row Cache.• Level 2/3 - Block Cache RAM/SSD.• YCSB benchmark results• Upcoming features in R1.5, 2.0, 3.0.• Q&A
HBase
• Still lacks some original BigTable’s features.• Still not able to utilize efficiently all RAM. • No good mixed storage (SSD/HDD) support. • Single Level Caching only. Simple. • HBase + Large JVM Heap (MemStore) = ?
BigBase
• Adds Row Cache and block cache compression.• Utilizes efficiently all RAM (TBs). • Supports mixed storage (SSD/HDD). • Has Multi Level Caching. Not that simple. • Will move MemStore off heap in R2.
BigBase History
Koda (2010)• Koda - Java off heap object cache, similar to
Terracotta’s BigMemory.• Delivers 4x times more transactions …• 10x times better latencies than BigMemory 4.• Compression (Snappy, LZ4, LZ4HC, Deflate).• Disk persistence and periodic cache snapshots.• Tested up to 240GB.
Karma (2011-12)• Karma - Java off heap BTree implementation to support
fast in memory queries.• Supports extra large heaps, 100s millions – billions
objects.• Stores 300M objects in less than 10G of RAM.• Block Compression.• Tested up to 240GB.• Off Heap MemStore in R2.
Yamm (2013)• Yet Another Memory Manager.– Pure 100% Java memory allocator.– Replaced jemalloc in Koda. – Now Koda is 100% Java.– Karma is the next (still on jemalloc).– Similar to memcached slab allocator.
• BigBase project started (Summer 2013).
BigBase Architecture
MLC – Multi-Level Caching
HBase 0.94
Disk
JVM
RA
M
LRUBlockCache
MLC – Multi-Level Caching
HBase 0.94
Disk
JVM
RA
M
LRUBlockCache
HBase 0.96
Disk
JVM
RA
M
Bucket cache
One level of caching : • RAM (L2)
MLC – Multi-Level Caching
HBase 0.94
Disk
JVM
RA
M
LRUBlockCache
HBase 0.96
Bucket cache
JVM
RA
M
One level of caching : • RAM (L2)• Or DISK (L3)
MLC – Multi-Level Caching
HBase 0.94
Disk
JVM
RA
M
LRUBlockCache
HBase 0.96
Disk
JVM
RA
M
Bucket cache
BigBase 1.0
Block Cache L3SSD
JVM
RA
M
Row Cache L1
Block Cache L2
MLC – Multi-Level Caching
HBase 0.94
Disk
JVM
RA
M
LRUBlockCache
HBase 0.96
Disk
JVM
RA
M
Bucket cache
BigBase 1.0
JVM
RA
M
Row Cache L1
Block Cache L2
BlockCache L3Network
MLC – Multi-Level Caching
HBase 0.94
Disk
JVM
RA
M
LRUBlockCache
HBase 0.96
Disk
JVM
RA
M
Bucket cache
BigBase 1.0
JVM
RA
M
Row Cache L1
Block Cache L2
BlockCache L3memcached
MLC – Multi-Level Caching
HBase 0.94
Disk
JVM
RA
M
LRUBlockCache
HBase 0.96
Disk
JVM
RA
M
Bucket cache
BigBase 1.0
JVM
RA
M
Row Cache L1
Block Cache L2
BlockCache L3DynamoDB
BigBase Row Cache (L1)
Where is BigTable’s Scan Cache?
• Scan Cache caches hot rows data. • Complimentary to Block Cache.• Still missing in HBase (as of 0.98). • It’s very hard to implement in Java (off heap).• Max GC pause is ~ 0.5-2 sec per 1GB of heap• G1 GC in Java 7 does not resolve the problem.• We call it Row Cache in BigBase.
Row Cache vs. Block Cache
HFile Block HFile BlockHFile BlockHFile BlockHFile Block
Row Cache vs. Block Cache
Row Cache vs. Block Cache
BLOCK CACHE
ROW CACHE
Row Cache vs. Block Cache
ROW CACHE
BLOCK CACHE
Row Cache vs. Block Cache
ROW CACHE
BLOCK CACHE
BigBase Row Cache
• Off Heap Scan Cache for HBase.• Cache size: 100’s of GBs to TBs. • Eviction policies: LRU, LFU, FIFO,
Random. • Pure 100% - compatible Java. • Sub-millisecond latencies, zero GC.• Implemented as RegionObserver
coprocessor.
Row Cache
YAMM Codecs Kryo SerDe
KODA
BigBase Row Cache
• Read through cache. • It caches rowkey:CF. • Invalidates key on every mutation.• Can be enabled/disabled per table and
per table:CF.• New ROWCACHE attribute.• Best for small rows (< block size)
Row Cache
YAMM Codecs Kryo SerDe
KODA
Performance-Scalability
• GET (small rows < 100 bytes): 175K operations per sec per one Region Server (from cache).
• MULTI-GET (small rows < 100 bytes): > 1M records per second (network limited) per one Region Server.
• LATENCY : 99% < 1ms (for GETs) with 100K ops.• Vertical scalability: tested up to 240GB (the maximum available
in Amazon EC2).• Horizontal scalability: limited by HBase scalability. • No more memcached farms in front of HBase clusters.
BigBase Block Cache (L2, L3)
What is wrong with Bucket Cache?
Scalability LIMITED
Multi-Level Caching (MLC) NOT SUPPORTED
Persistence (‘offheap’ mode) NOT SUPPORTED
Low latency apps NOT SUPPORTED
SSD friendliness (‘file’ mode) NOT FRIENDLY
Compression NOT SUPPORTED
What is wrong with Bucket Cache?
Scalability LIMITED
Multi-Level Caching (MLC) NOT SUPPORTED
Persistence (‘offheap’ mode) NOT SUPPORTED
Low latency apps NOT SUPPORTED
SSD friendliness (‘file’ mode) NOT FRIENDLY
Compression NOT SUPPORTED
What is wrong with Bucket Cache?
Scalability LIMITED
Multi-Level Caching (MLC) NOT SUPPORTED
Persistence (‘offheap’ mode) NOT SUPPORTED
Low latency apps NOT SUPPORTED
SSD friendliness (‘file’ mode) NOT FRIENDLY
Compression NOT SUPPORTED
What is wrong with Bucket Cache?
Scalability LIMITED
Multi-Level Caching (MLC) NOT SUPPORTED
Persistence (‘offheap’ mode) NOT SUPPORTED
Low latency apps ?
SSD friendliness (‘file’ mode) NOT FRIENDLY
Compression NOT SUPPORTED
What is wrong with Bucket Cache?
Scalability LIMITED
Multi-Level Caching (MLC) NOT SUPPORTED
Persistence (‘offheap’ mode) NOT SUPPORTED
Low latency apps NOT SUPPORTED
SSD friendliness (‘file’ mode) NOT FRIENDLY
Compression NOT SUPPORTED
What is wrong with Bucket Cache?
Scalability LIMITED
Multi-Level Caching (MLC) NOT SUPPORTED
Persistence (‘offheap’ mode) NOT SUPPORTED
Low latency apps NOT SUPPORTED
SSD friendliness (‘file’ mode) NOT FRIENDLY
Compression NOT SUPPORTED
Here comes BigBase
Scalability HIGH
Multi-Level Caching (MLC) SUPPORTED
Persistence (‘offheap’ mode) SUPPORTED
Low latency apps SUPPORTED
SSD friendliness (‘file’ mode) SSD-FRIENDLY
Compression SNAPPY, LZ4, LZHC, DEFLATE
Here comes BigBase
Scalability HIGH
Multi-Level Caching (MLC) SUPPORTED
Persistence (‘offheap’ mode) SUPPORTED
Low latency apps SUPPORTED
SSD friendliness (‘file’ mode) SSD-FRIENDLY
Compression SNAPPY, LZ4, LZHC, DEFLATE
Here comes BigBase
Scalability HIGH
Multi-Level Caching (MLC) SUPPORTED
Persistence (‘offheap’ mode) SUPPORTED
Low latency apps SUPPORTED
SSD friendliness (‘file’ mode) SSD-FRIENDLY
Compression SNAPPY, LZ4, LZHC, DEFLATE
Here comes BigBase
Scalability HIGH
Multi-Level Caching (MLC) SUPPORTED
Persistence (‘offheap’ mode) SUPPORTED
Low latency apps SUPPORTED
SSD friendliness (‘file’ mode) SSD-FRIENDLY
Compression SNAPPY, LZ4, LZHC, DEFLATE
Here comes BigBase
Scalability HIGH
Multi-Level Caching (MLC) SUPPORTED
Persistence (‘offheap’ mode) SUPPORTED
Low latency apps SUPPORTED
SSD friendliness (‘file’ mode) SSD-FRIENDLY
Compression SNAPPY, LZ4, LZHC, DEFLATE
Here comes BigBase
Scalability HIGH
Multi-Level Caching (MLC) SUPPORTED
Persistence (‘offheap’ mode) SUPPORTED
Low latency apps SUPPORTED
SSD friendliness (‘file’ mode) SSD-FRIENDLY
Compression SNAPPY, LZ4, LZHC, DEFLATE
Wait, there are more …
Scalability HIGHMulti-Level Caching (MLC) SUPPORTEDPersistence (‘offheap’ mode) SUPPORTEDLow latency apps SUPPORTEDSSD friendliness (‘file’ mode) SSD-FRIENDLYCompression SNAPPY, LZ4, LZHC, DEFLATENon disk–based L3 cache SUPPORTEDRAM Cache optimization IBCO
Wait, there are more …
Scalability HIGHMulti-Level Caching (MLC) SUPPORTEDPersistence (‘offheap’ mode) SUPPORTEDLow latency apps SUPPORTEDSSD friendliness (‘file’ mode) SSD-FRIENDLYCompression SNAPPY, LZ4, LZHC, DEFLATENon disk–based L3 cache SUPPORTEDRAM Cache optimization IBCO
BigBase 1.0 vs. HBase 0.98
BigBase HBase 0.98
Row Cache (L1) YES NO
Block Cache RAM (L2) YES (fully off heap) YES (partially off heap)
Block Cache (L3) DISK YES (SSD- friendly) YES (not SSD – friendly)
Block Cache (L3) NON DISK YES NO
Compression YES NO
RAM Cache persistence YES (both L1 and L2) NO
Low Latency optimized YES NO
MLC support YES (L1, L2, L3) NO (either L2 or L3)
Scalability HIGH MEDIUM (limited by JVM heap)
YCSB Benchmark
Test setup (AWS)
• HBase 0.94.15 – RS: 11.5GB heap (6GB LruBlockCache on heap); Master: 4GB heap.
• Clients: 5 (30 threads each), collocated with Region Servers.
• Data sets: 100M and 200M. 120GB / 240GB approximately. Only 25% fits in a cache.
• Workloads: 100% read (read100, read200, hotspot100), 100% scan (scan100, scan200) –zipfian.
• YCSB 0.1.4 (modified to generate compressible data). We generated compressible data (with factor of 2.5x) only for scan workloads to evaluate effect of compression in BigBase block cache implementation.
• Common – Whirr 0.8.2; 1 (Master + Zk) + 5 RS; m1.xlarge: 15GB RAM, 4 vCPU, 4x420 HDD
• BigBase 1.0 (0.94.15) – RS: 4GB heap (6GB off heap cache); Master: 4GB heap.
• HBase 0.96.2 – RS: 4GB heap (6GB Bucket Cache off heap); Master: 4GB heap.
Test setup (AWS)
• HBase 0.94.15 – RS: 11.5GB heap (6GB LruBlockCache on heap); Master: 4GB heap.
• Clients: 5 (30 threads each), collocated with Region Servers.
• Data sets: 100M and 200M. 120GB / 240GB approximately. Only 25% fits in a cache.
• Workloads: 100% read (read100, read200, hotspot100), 100% scan (scan100, scan200) –zipfian.
• YCSB 0.1.4 (modified to generate compressible data). We generated compressible data (with factor of 2.5x) only for scan workloads to evaluate effect of compression in BigBase block cache implementation.
• Common – Whirr 0.8.2; 1 (Master + Zk) + 5 RS; m1.xlarge: 15GB RAM, 4 vCPU, 4x420 HDD
• BigBase 1.0 (0.94.15) – RS: 4GB heap (6GB off heap cache); Master: 4GB heap.
• HBase 0.96.2 – RS: 4GB heap (6GB Bucket Cache off heap); Master: 4GB heap.
Benchmark results (RPS)
BigBase R1.0 HBase 0.96.2 HBase 0.94.150
2000
4000
6000
8000
10000
12000
14000
16000
11405
6123 55536265
4086 3850
15150
3512 28553224
1500709820 434 228
read100read200hotspot100scan100scan200
Average latency (ms)
BigBase R1.0 HBase 0.96.2 HBase 0.94.150
100
200
300
400
500
600
700
800
13 24 2723 36 3910 44 5248102
223187
375
700
read100read200hotspot100scan100scan200
95% latency (ms)
BigBase R1.0 HBase 0.96.2 HBase 0.94.150
100200300400500600700800900
1000
51 91 10088 124 13838
152197175
405
950
729
read100read200hotspot100scan100scan200
99% latency (ms)
BigBase R1.0 HBase 0.96.2 HBase 0.94.150
100
200
300
400
500
600
700
800
900
133190 213225
304 338
111
554632
367
811
read100read200hotspot100scan100scan200
YCSB 100% Read
BigBase R1.0 HBase 0.94.150
5001000150020002500300035004000 3621
1308
2281
11111253770
Per Server
50M 100M 200M
• 50M = 2.77X• 100M = 2.05X• 200M = 1.63X• 50M = 40% fits cache• 100M = 20% fits cache• 200M = 10% fits cache• What is the maximum?
YCSB 100% Read
BigBase R1.0 HBase 0.94.150
5001000150020002500300035004000 3621
1308
2281
11111253770
Per Server
50M 100M 200M
• 50M = 2.77X• 100M = 2.05X• 200M = 1.63X• 50M = 40% fits cache• 100M = 20% fits cache• 200M = 10% fits cache• What is the maximum?• ~ 75X (hotspot 2.5/100)• 56K (BB) vs. 750 (HBase)• 100% in cache
All data in cache
• Setup: BigBase 1.0, 48G RAM, (8/16) CPU cores – 5 nodes (1+ 4)
• Data set: 200M (300GB) • Test: Read 100%, hotspot• YCSB 0.1.4 – 4 clients• 40 threads – 100K• 100 threads – 168K• 200 threads – 224K• 400 threads - 262K
100,000 168,000 224,000 262,000
99%
1 2 3 7
95%
1 1 2 3
avg 0.4 0.6 0.9 1.5
0.52.54.56.5
Hotspot (2.5/100 – 200M data)La
tenc
y (m
s)
All data in cache
• Setup: BigBase 1.0, 48G RAM, (8/16) CPU cores – 5 nodes (1+ 4)
• Data set: 200M (300GB) • Test: Read 100%, hotspot• YCSB 0.1.4 – 4 clients• 40 threads – 100K• 100 threads – 168K• 200 threads – 224K• 400 threads - 262K
100,000 168,000 224,000 262,000
99%
1 2 3 7
95%
1 1 2 3
avg 0.4 0.6 0.9 1.5
0.52.54.56.5
Hotspot (2.5/100 – 200M data)La
tenc
y (m
s)
100K ops: 99% < 1ms
What is next?
• Release 1.1 (2014 Q2)– Support HBase 0.96, 0.98, trunk– Fully tested L3 cache (SSD)
• Release 1.5 (2014 Q3)– YAMM: memory allocator compacting mode .– Integration with Hadoop metrics.– Row Cache: merge rows on update (good for counters).– Block Cache: new eviction policy (LRU-2Q).– File read posix_fadvise ( bypass OS page cache).– Row Cache: make it available for server-side apps
What is next?
• Release 2.0 (2014 Q3)– HBASE-5263: Preserving cache data on compaction– Cache data blocks on memstore flush (configurable). – HBASE-10648: Pluggable Memstore. Off heap implementation,
based on Karma (off heap BTree lib).• Release 3.0 (2014 Q4)
– Real Scan Cache – caches results of Scan operations on immutable store files.
– Scan Cache integration with Phoenix and with other 3rd party libs provided rich query API for HBase.
Download/Install/Uninstall• Download BigBase 1.0 from www.bigbase.org• Installation/upgrade takes 10-20 minutes• Beatification operator EM(*) is invertible:
HBase = EM-1(BigBase) (the same 10-20 min)
Q & A
Vladimir RodionovHadoop/HBase architectFounder of BigBase.org
HBase: Extreme makeoverFeatures & Internal Track