Team webapp7 https://lb.webapp.ku.whs.in.th 

Why HBase? ● MapReduce is slowwwwwww ● Takes 40 minutes to import data using MapReduce to HBase ● Too many small files are weak point of Hadoop's MapReduce 

HTable structure Key timestamp p:f p:p a 1464292227911 rmlav10.txt 613 e 1464292227911 12370-8.txt 387 

The problem with HBase ● HBase + Hadoop consumes too many memory! ● Solution: 1GB swap file 

HBase architecture Load Balancer momo Web Server konoha Web Server ene Zookeeper Master Hadoop Namenode ayano Zookeeper Secondary Master Hadoop Datanode kido Zookeeper Quorum Hadoop Datanode seto nginx Rapidoid (Java) Rapidoid (Java) 

Benchmark 

Can we go faster? 

The "fast" implementation https://fast.lb.webapp.ku.whs.in.th 'n reckless 

Benchmark ● Run on load balancer ● wrk benchmark ○ 30s test ○ 500 connections ● HBase and MapFile has some timeouts Query: Corrected EDITIONS of our eBooks replace the old file and take over the old filename and etext number. The replaced older file is renamed. EBooks posted prior to November 2003, with eBook numbers BELOW #10000, are filed in directories based on their release date. If you want to download any of these eBooks directly, rather than using the regular search system you may utilize the following addresses and just download by the etext year. 2.87x 

Fast search data structure 

HashMap! 

The size problem 1 2 3 4 5 Key Value Key Value Key xxHash 1 

The size problem 1 2 3 4 5 The Common Edition is the result of a decade long study of the New Testament and numerous English translations in the modern church. The goals for this edition are: { "file": "Introduction_and_Copyright.txt", "paragraph": 2 } 

"When you see something big, make it small enough" — JL ไมไดกลาวไว 1 2 3 4 5 1234561561865 { "file": "Introduction_and_Copyright.txt", "paragraph": 2 } 

The size problem 1 2 3 4 5 1234561561865 0001 0010 1 Introduction_and_Copyright.txt struct { int file_id; int paragraph_id; } 

Database size 

HashMap implementation ● Initially wrote a custom memory-mapped HashMap in C ● Due to memory limitation of mmap, database size is limited to available memory ● Changed to LMDB from OpenLDAP (via Java binding) ● LMDB database size limit is set to 1GB ● May not scale out really well – Risk of hash collision 

รบรอยชนะรอย ยังไมดีพอ ชนะโดยไมตองรบ แมเพียงครั้งเดียว ยอมถือวาเลิศจบแดน – Sun Tzu 

Bloom filter Hello world!! 3 4 xxhash murmur3 

Bloom filter Hello world!! 3 4 xxhash murmur3 

Bloom filter Yes! 1 3 xxhash murmur3 

Bloom filter Noooo 3 4 xxhash murmur3 

Bloom filter ● 200kB Bloom filter is used = 1,638,400 slots! ● Using a custom Java implementation, every bits are used! ● Bloom filter is preloaded in memory of the server process when it starts ● Planned to send Bloom filter to run on the clientside ○ Different data type limit ○ Complex client side code 

Did the Bloom filter helped? ● https://en.wikipedia. org/wiki/Opinion_polling_for_the_2015_United_K ingdom_general_election ○ Content Length 289,085 ● Romano Lavo-Lil (rmlav10.txt) ○ Size 288,195 ● HBase implementation also use the same Bloom filter 

LMDB architecture Load Balancer momo Web Server konoha Web Server ene 

HDFS Edition ● http://hdfs.lb.webapp.ku.whs.in.th (No HTTPS) ● Using MapReduce job to generate MapFile ● Just wrote this morning 

Thank you https://github.com/whsatku/lmdbsearch