A Train of Thought About Growing and Scalability Bumping up Startup Business with Apache Cassandra Eiti Kimura, Software Engineer @
#Cassandra Summit 2014 - A Train of Thoughts About Growing and Scalability
Jul 14, 2015
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
A Train of Thought About
Growing and Scalability
Bumping up Startup Business with Apache Cassandra
Eiti Kimura, Software Engineer @
Movile is the industry leader for development of mobile
content and commerce platforms in Latin America. With
products for mobile phones, smartphones and tablets.
Games, on-line education, entertainment apps for adults and
kids and many options for buying with confidence and
comfort. All of that comes to you through Movile.
For companies, Movile delivers complete products,
integrating transactions in M-Payments and content
distribution, fast and with quality.
Subscription and Billing Platform a.k.a SBS
• It is a service API
• responsible to manage users subscriptions
• charge users in carriers
• renew subscriptions
• “can not” stop anyway
• should be as performatic as possible
Some platform numbers
Renewal Engine: ~ 52,1
million of billing tries a
day
• about 603 request/s
• 1,5 billion billing tries
per month
50 million
subscriptions
~ 50 request/s
Operations:
★ subscribe
★ cancel
★profile
Data DistributionSubscriptions by Country
1%2%
4%
25%
68%
Others
Colombia
Argentina
Mexico
Brazil
Platform Architecture
“There isn’t just one way to state a system’s architecture; rather, there are
multiple architectures in a system, and the view of what is architecturally
significant is one that can change over a system’s lifetime.” - Patterns of
Enterprise Application Architecture
Martin Fowler
Basic Architecture Design
• scalabe
• high availability
• high performance
Very High Usage
• veryyyyy slow... system response
• overall throughput decreased
• low availability, single point of
failure
• Even worse than stopping is to only
work sometimes
Improved Distributed Design
A Cassandra Based Solution
• the operations are
distributed across the
nodes
• we achieved linear
scalability
Improved Distributed Design
A Cassandra Based Solution
• the performance issues
were solved
• the availability has improved
• there is no longer a single
point of failure
C* Data Modeling
• Dernormalization: Writes are cheap, reads are expensive, so insert data in every arrangement that
you need to read
• Don't be afraid of denormalization
• There are different ways to model your solution, there is no right or wrong way
• plan your queries and how you need to get the information before modeling. Use it as driver for
modeling decisions
Data Model V1
CREATE TABLE subscription (
subscription-id text PRIMARY KEY,
phone-number text,
config-id int,
…
enabled boolean,
creation-date timestamp
);
CREATE TABLE user_subscriptions (
phone-number text,
subscription-id text,
PRIMARY KEY (phone-number, subscription-id)
);
Data Model V1
user_subscriptions
phone-number subscription-id
551900001212 subs-093123
551911114567 subs-002202
551911114567 subs-002203
551911114567 subs-002204
subscriptions
subscription-id phone-number config-id . . . enabled creation-date
subs-093123 551900001212 342 . . . true 2013-08-01
subs-002202 551911114567 567 . . . false 2014-06-27
subs-002203 551911114567 678 . . . true 2014-07-05
subs-002204 551911114567 654 . . . true 2014-08-07
Data Model V1 – Quering Profile
user_subscriptions
phone-number subscription-id
551900001212 subs-093123
551911114567 subs-002202
551911114567 subs-002203
551911114567 subs-002204
#cql> _
1st step
• check the index table to get the ids of
subscriptions for a given user
Data Model V1 – Quering Profile
user_subscriptions
phone-number subscription-id
551900001212 subs-093123
551911114567 subs-002202
551911114567 subs-002203
551911114567 subs-002204
#cql> SELECT * FROM user_subscriptions WHERE phone-number = 551911114567;
551911114567 subs-002202
551911114567 subs-002203
551911114567 subs-002204
Data Model V1 – Quering Profile
#cql> _
2nd step
• query all the user’s subscriptions by its id
551911114567 subs-002202
551911114567 subs-002203
551911114567 subs-002204
Data Model V1 – Quering Profile
#cql> SELECT * FROM subscriptions WHERE subscription-id = ‘subs-002204’;
#cql> SELECT * FROM subscriptions WHERE subscription-id = ‘subs-002203’;
#cql> SELECT * FROM subscriptions WHERE subscription-id = ‘subs-002202’;
subscriptions
subscription-id phone-number config-id . . . enabled creation-date
subs-093123 551900001212 342 . . . true 2013-08-01
subs-002202 551911114567 567 . . . false 2014-06-27
subs-002203 551911114567 678 . . . true 2014-07-05
subs-002204 551911114567 654 . . . true 2014-08-07
Data Model V2
CREATE TABLE subscription (
phone-number text,
subscription-id text,
serialized blob,
PRIMARY KEY(phone-number, subscription-id)
);
Data Model V2
subscriptions
phone-number subscription-id serialized-data
551900001212 subs-093123 array [1,1,0,1,1,1,1,0,0,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,0]
551911114567 subs-002202 array [0,1,0,1,1,0,1,1,0,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,0]
551911114567 subs-002203 array [0,1,0,0,1,1,1,0,0,1,0,10,1,1,0,1,1,0,1,1,1,1,1,1,1,0]
551911114567 subs-002203 array [1,0,0,1,1,1,1,0,1,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,1]
542154231121 subs-320012 array [1,1,1,1,1,0,1,0,1,0,0,10,1,0,0,1,1,0,1,1,1,0,0,1,0,1]
Data Model V2 – Quering Profile
#cql> SELECT * FROM subscriptions WHERE phone-number = ‘551911114567’;
subscriptions
phone-number subscription-id serialized-data
551900001212 subs-093123 array [1,1,0,1,1,1,1,0,0,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,0]
551911114567 subs-002202 array [0,1,0,1,1,0,1,1,0,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,0]
551911114567 subs-002203 array [0,1,0,0,1,1,1,0,0,1,0,10,1,1,0,1,1,0,1,1,1,1,1,1,1,0]
551911114567 subs-002204 array [1,0,0,1,1,1,1,0,1,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,1]
542154231121 subs-320012 array [1,1,1,1,1,0,1,0,1,0,0,10,1,0,0,1,1,0,1,1,1,0,0,1,0,1]
551911114567 subs-002202 array [0,1,0,1,1,0,1,1,0,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,0]
551911114567 subs-002203 array [0,1,0,0,1,1,1,0,0,1,0,10,1,1,0,1,1,0,1,1,1,1,1,1,1,0]
551911114567 subs-002204 array [1,0,0,1,1,1,1,0,1,0,0,10,1,1,0,1,1,0,1,1,1,0,0,0,1,1]
Storage Strategy
• we tried some various ways to store
information
• it optimizes network traffic as well
0
500
1000
1500
2000
2500
1O
bje
ct
siz
e in
byte
s
Types of object representation
Object Representation
XML Java Byte Array JSON protobuff
Database volume
• the database size, decreases
considerably
• less data to handle, more
performance
0.00 20.00 40.00 60.00 80.00 100.00
XML
Java Byte Array
JSON
protobuff
Data Size in GB
Sto
rag
e S
trate
gy
Subscription Data Volume
C* New Data Model
• performance increased significantly
• reduced complexity: from 2 tables to 1, simpler, lighter
• reduced number of remote calls
• V1
• 1 query to the index table
• X queries (one per index returned)
• V2
• 1 query brings all data
• data volume reduced
Cassandra Cluster Configuration
• Geographically distributed
• 2 data centers in São Paulo Brazil
Ring Topology
Datacenter: DC1===============Status=Up/Down|/ State=Normal/Leaving/Joining/Moving-- Address Load Tokens Owns (effective) Host ID RackUN 200.xxx.xxx.73 29.58 GB 256 76,7% b9f890b6-6137-4359-90c2-74f87ce1676d RAC1UN 200.xxx.xxx.72 29.8 GB 256 74,5% ec7fa873-edd9-4cb9-938d-60f1c9b8f742 RAC1UN 200.xxx.xxx.71 30.76 GB 256 76,1% 1091799e-0617-42dd-a396-363f10c03295 RAC1UN 200.xxx.xxx.74 26.68 GB 256 72,7% 984b848b-0ecb-4db3-a1fe-c9b088c295f6 RAC1
Datacenter: DC2===============Status=Up/Down|/ State=Normal/Leaving/Joining/Moving-- Address Load Tokens Owns (effective) Host ID RackUN 200.xxx.xxx.72 28.99 GB 256 100,0% f9b820d6-111f-4a3a-af6c-39d0e8e88084 RAC1UN 200.xxx.xxx.71 30.36 GB 256 100,0% 120939bd-a6b4-4d88-b2cf-dbf79d93181c RAC1UN 200.xxx.xxx.74 27.93 GB 256 100,0% c821b8f7-2224-4512-8a0e-0371460d900e RAC1
nodetool status
Hardware Infrastructure v1.0
• Centos 5.9
• 2x Intel(R) Xeon(R) CPU E5606 @ 2.13GHz (4 cores)
• 24GB / 32GB RAM
• 1x SATA 500gb (OS)
• 1x SSD CSSD-F120GB2 (data)
• Apache Cassandra v1.0.6
4 Servers
Hardware Infrastructure v2.0
• 2 Intel (R) Xeon (R) CPU @3.1GHz
• 128 GB of total RAM Memory per Server
• Running Cent OS 6.5
• 32 GB of RAM per VM
• 1 Intel (R) Xeon (R) CPU @3.1GHz
• 2 SSD Disks Model : CSSD-F120GBGT
• Configured as RAID0
• Apache Cassandra 1.2.13
6 Servers
6 Virtual Machines (one per physical server)
VMs
Keyspace
Keyspace: SBSPlatform:Replication Strategy: org.apache.cassandra.locator.NetworkTopologyStrategy
Options: [DC2:3, DC1:3]
cassandra-cli : describe
Column Families:ColumnFamily: subscriptionColumnFamily: delivery_ticketColumnFamily: hard_limit_controlColumnFamily: hard_limit_rulesColumnFamily: idx_config_subscColumnFamily: user_directives
Column Family Status
Column Family: subscriptionSpace used (total): 13499012297Number of Keys (estimate): 46.369.536Read Count: 5599788263 / Read Latency: 0,497 ms.Write Count: 5212858995 / Write Latency: 0,017
ms.Compacted row mean size: 576
./nodetool cfstats SBSPlatform
Column Family: hard_limit_controlSpace used (total): 7812531598Number of Keys (estimate): 44.785.024Read Count: 3987345295 / Read Latency: 0,509 ms.Write Count: 11646786043 / Write Latency: 0,021 ms.Compacted row mean size: 188
Overall cluster response time
Node 1 - : 200.xxx.xxx.71load_avg: 0.39write_latency(us): 900.8read_latency(us): 553.6
Node 2 - : 200.xxx.xxx.72load_avg: 0.51write_latency(us): 874.1read_latency(us): 620.5
Node 3 - : 200.xxx.xxx.74load_avg: 0.35write_latency(us): 834.87read_latency(us): 515.6
Node 4 - : 200.xxx.xxx.73load_avg: 0.35write_latency(us): 900.87read_latency(us): 700.6
Node 1 - : 200.xxx.xxx.71load_avg: 0.63write_latency(us): 806.3read_latency(us): 882.3
Node 2 - : 200.xxx.xxx.72load_avg: 0.37write_latency(us): 802.8read_latency(us): 969.0
Node 3 - : 200.xxx.xxx.74load_avg: 0.62write_latency(us): 965.7read_latency(us): 887.43
Now: 2014-08-30 14:49:15Total Reads/second: 13262Total Writes/second: 9529
DATACENTER 1DATACENTER 2
O.S. and Software Customizations
• Java 1.7 + JNA
• Disable Swap
• NTP server in all servers
According to the Cassandra Docs the Recommended Settings for Production
O.S. - limits.conf
# number of open filesroot soft nofile 100000root hard nofile 100000* soft nofile 100000* hard nofile 100000
# allocated memoryroot soft memlock unlimited root hard memlock unlimited * soft memlock unlimited * hard memlock unlimited
# addressing (virtual memory)root soft as unlimited root hard as unlimited * soft as unlimited * hard as unlimited
# number of open processesroot soft nproc unlimitedroot hard nproc unlimited* soft nproc unlimited* hard nproc unlimited
Daily Cluster Operations
Total Reads
Total Writes
> 1 billion
Conclusion: Why Cassandra?
• Good performance for Reads
• Excellent performance for Writes
• Read and Write throughput highly scalable (linear)
• Supports GEO distributed information
• Fault Tolerant
• Tunable consistency per client
• FOSS (Free and Open Source Software) + Support
Related Documents
