Presentation is loading. Please wait.

Presentation is loading. Please wait.

University of Illinois at Urbana-Champaign

Similar presentations


Presentation on theme: "University of Illinois at Urbana-Champaign"— Presentation transcript:

1 University of Illinois at Urbana-Champaign
Parqua Online Reconfigurations in Ring-Based NoSQL Systems Yosub Shin, Mainak Ghosh, Indranil Gupta University of Illinois at Urbana-Champaign DPRG:

2 Ring-Based Key-value Stores
Growing segment in Industry (21% CAGR, $3.4 B by 2020) Replacing traditional databases Fast and highly available Apache Cassandra (Facebook), Voldemort (LinkedIn), Riak (Basho), Dynamo (Amazon)… Used widely Amazon shopping carts (Dynamo) Netflix (Cassandra) 2

3 Reconfiguration Operations
Major pain point: Reconfiguration operations Change that affects a lot of data in tables at once Examples Change key, e.g., SSN  Employee ID [MongoDB JIRA Server-4000] Scale out/in Today’s solution: export DB, shut down DB, change configuration, re-import and re-start DB 3

4 Parqua System Allows ring-based key-value stores to perform reconfiguration in online manner Answering reads, writes and queries normally While transferring data in the background among servers and committing 4

5 Background: Apache Cassandra – Virtual Ring
5

6 Reconfiguration: Parqua Phases
I. Prepare Phase Create new table Column family (CF) Old CF continues to serve reads/writes 6

7 Reconfiguration: Parqua Phases
II. Execute Phase Each node Hashes its data Sends it to new node in new configuration Appropriate replica 7

8 Reconfiguration: Parqua Phases
III. Commit Phase Atomically swap between old and new configs Schemas SSTables Writes locked briefly 8

9 Reconfiguration: Parqua Phases
IV. Recovery Phase Nodes catch up with R/W received during Execute phase Activate new config. 9

10 Experiments Parqua: Maintains high availability under reconfiguration
multiple 9’s of availability Has minimal impact on latency 50th percentile about the same Tail slightly longer Scales with database size and cluster size Read (%) Write (%) Read only 99.17 Uniform 99.27 99.01 Latest 96.07 98.92 Zipfian 99.02 10 DPRG:

11 Backup Slides 11

12 Experiments Effect on Read Latency 12

13 Experiments Scalability (Database size)

14 Experiments Scalability (Cluster size) 14


Download ppt "University of Illinois at Urbana-Champaign"

Similar presentations


Ads by Google