Building High Performance Apps with In-memory Data

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved
Pop-up Loft
Building High Performance Apps with In-Memory Data
Mikhail Prudnikov
prudniko@amazon.com
Sr Solutions Architect

© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
In-memory key-value store supporting
• Redis 3.2.10
• Memcached 1.4.34
High-performance
Fully managed; zero admin
Highly available and reliable
Hardened by Amazon
Amazon
ElastiCache

Powerful
~200 commands + Lua scripting
In-memory data structure server
Utility data structures
Strings, lists, hashes, sets, sorted sets,
bitmaps & HyperLogLogs
Simple
Atomic operations
supports transactions
Ridiculously fast!
<1ms latency for most commands
Highly available
replication
Persistence
Open source
Redis overview

S5
S1
S2
S4 S3
Client
§ 16384 hash slots per cluster
§ Slot for a key is CRC16(key) mod 16384
§ Slots are distributed across the cluster into shards
§ Developers must use a Redis cluster aware client
§ Clients are redirected to the correct shard
§ Smart clients store a map
Shard S1 = slots 0–3276
Shard S5 = slots 13107–16383
Redis cluster: automatic client-side sharding

reads/
writes
reads
AZ1
AZ2
reads
search
reads
search
clients
c a c h e c l u s t e r
r e l a t i o n a l
d a t a
Healthy

reads/
writes
reads
AZ1
AZ2
reads
search
reads
search
clients
r e l a t i o n a l
d a t a
Heavy
pressure

AWS Lambda
3 Shards
…
var params = {
ApplyImmediately: true,
NodeGroupCount: 5,
ReplicationGroupId: ‘rep-group-id’,
… }
elasticache.modifyReplicationGroupShardConfiguration(params, function(err, data) {
if (err) console.log(err, err.stack);
else console.log(data);
}); …
5 Shards
MEMORY
HIGH!
Amazon
CloudWatch
Cluster Resized
AWS SNS
Online Re-Sharding—CW alarm triggered

reads/
writes
reads
AZ1
AZ2
reads
search
reads
search
clients
r e l a t i o n a l
d a t a
Healthy—
auto scaled
out

Session
management
Database caching APIs
(HTTP responses)
IOT
Streaming data
analytics
(Filtering/aggregation)
Pub/sub
Social media
(Sentiment analysis)
Standalone
database
(Metadata store)
Leaderboards
Usage Patterns

Caching
Clients
Amazon
ElastiCache
Redis
Amazon
DynamoDB
Elastic Load
Balancing
Amazon
EC2
Amazon
RDS
write-through
reads/
writes
DDB streams
mysql.lambda_async
reads/
writes
Amazon
S3
reads/writes
Object data
Unstructured data
Relational data

Caching NoSQL
Amazon
EC2 reads/
writes
reads
MongoDB
Cluster
Cassandra
Cluster
ü Smaller NoSQL DB clusters needed = lower costs
ü Faster data retrieval = better performance
Elasticsearch
Cluster
Clients

Amazon
EC2
reads/
writes
Amazon
ElastiCache
Redis
reads
MongoDB
Cluster
DBObject doc = collection.findOne();
Cache serialized DBObject in Redis (good)
Cache rows in Redis hash (faster/more efficient)
Cassandra
Cluster
Amazon
ElastiCache
Redis
Amazon
EC2
reads/
writes
reads
ResultSet rs = session.execute(stmt);
Cache serialized ResultSet in Redis (good)
Cache rows in Redis hash (faster/more efficient)
ü Smaller NoSQL DB clusters needed = lower costs
ü Faster data retrieval = better performance
Caching NoSQL databases with Amazon ElastiCache

Amazon
Kinesis
Analytics
Amazon
Kinesis
Streams
Amazon
Kinesis
Streams
Amazon
ElastiCache
(Redis)
cleansed
stream
Streaming data enrichment/processing
Datasources
raw
stream
Subscribers
AWS Lambda function 1
Continual data
filtering/enrichment
Real-time
pub/sub
AWS Lambda function 2

Big data architectures using Redis
Amazon Kinesis
DataSources
AWS Lambda
Apache Storm
on EMR
Spark Streaming
on Amazon EMR
Amazon
Kinesis app
Amazon
EC2
AWS IoT
Amazon
ElastiCache
Collect
Store
Process
Amazon
S3
Apache Kafka
AWS
Lambda
Custom
app
Spark on
Amazon
EMR
Analyze

Rules Engine
Amazon
ElastiCache
Redis
AWS
Lambda
Direct integration
LambdaSNS SQS
S3 KinesisDDB
AWS
IoT devices
AWS
IoT
Sensor store
IoT powered by ElastiCache

Mobile apps powered by ElastiCache
Amazon API
Gateway
AWS
Lambda
Amazon
ElastiCache
Redis
GEOADD
GEORADIUS
Search points of interest
Update points of interest
https://aws.amazon.com/blogs/database/amazon-elasticache-utilizing-redis-geospatial-capabilities/
Amazon
DynamoDB
DDB streams
Amazon
EC2

Ad tech powered by ElastiCache
Clients
Advertisers
https://aws.amazon.com/caching/database-caching/
Ad network
Ad slot
Consumer
Ad slot
publishers
Ad placement
(websites/apps)
Amazon
ElastiCache
Redis
<40 ms
Clickstream
(shopping
events)
User visits
page
Publisher
places ad slot
for auction
Ad network
calls for bidsBidders respond
with bids
Winners bid
ad displayed

Chat apps powered by ElastiCache
https://aws.amazon.com/blogs/database/amazon-elasticache-utilizing-redis-geospatial-capabilities/
Clients
Chat apps
Application Load
Balancer
WebSockets
Amazon
ElastiCache
Redis
PubSub
Serverpersistent
connections
Elastic
Beanstalk
SUBSCRIBE chat_channel:114
PUBLISH chat_channel:114 "Hello all"
>> ["message", "chat_channel:114", "Hello all"]
UNSUBSCRIBE chat_channel:114

• Very popular for gaming apps that need
uniqueness and ordering
• Easy with Redis sorted sets
ZADD "leaderboard" 1201 "Gollum”
ZADD "leaderboard" 963 "Sauron"
ZADD "leaderboard" 1092 "Bilbo"
ZADD "leaderboard" 1383 "Frodo”
ZREVRANGE "leaderboard" 0 -1
1) "Frodo"
2) "Gollum"
3) "Bilbo"
4) "Sauron”
ZREVRANK "leaderboard" "Sauron"
(integer) 3
Gaming—real-time leaderboards

Ex: throttling requests to an API
uses Redis counters
ELB
Externally
facing API
Reference: http://redis.io/commands/INCR
FUNCTION LIMIT_API_CALL(APIaccesskey)
limit = HGET(APIaccesskey, “limit”)
time = CURRENT_UNIX_TIME()
keyname = APIaccesskey + ":” + time
count = GET(keyname)
IF current != NULL && count > limit THEN
ERROR ”API request limit exceeded"
ELSE
MULTI
INCR(keyname)
EXPIRE(keyname,10)
EXEC
PERFORM_API_CALL()
END
Rate limiting

What data should I cache?

• R e f e r e n c e d a t a : p r o d u c t c a t e g o r i e s ( u s e r l o n g e r T T L s )
• R e f e r e n c e d a t a : p r o d u c t i m a g e s ( u s e r l o n g e r T T L s )
• R e f e r e n c e d a t a : p r o d u c t d e t a i l s ( u s e r l o n g e r T T L s )
• D y n a m i c : d a t a b a s e r e s u l t s e t s : ( u s e r s h o r t e r T T L s )
• D y n a m i c : A P I r e s p o n s e s ( u s e r s h o r t e r T T L s )
• A n y t h i n g t h a t i s c a c h e a b l e !
What data to cache

How do I cache my data?

Cache-aside—lazy loading
Cache-aside—lazy loading
Amazon
ElastiCache
Applications
Amazon
RDS
1
2
3
1) Check cache, if HIT return
2) If Cache MISS
3) Update cache

Cache-aside—write-through
Cache-aside—write-through
Amazon
ElastiCache
Applications
Amazon
RDS
1
2
1) Update primary DB
2) Update cache

Cache strategies
DB result set
Pro
When data retrieval logic is abstracted from the code consuming the ResultSet, caching the ROW can
be extremely effective and can be implemented against any RDBMS
Con
Data retrieval still requires extracting values from the ROW and does not further simplify data access, it
only reduces data-retrieval latency
SELECT * FROM x WHERE y
ID First_Name Last_Name City
123 Michael Labib Chicago
ResultSet object (ROW) Key: query, value: CRS as byte array

Cache strategies
JSON
Key: 123, Value: ‘{ “firstname”: “Michael”,
“lastname”: “Labib”,
“city”: “Chicago” } ‘
String firstName = rs.getString(First_Name)
Pro
Very easy to implement. Cache any desired database fields and values into a Redis string. For example,
store your retrieved data into a JSON object stored in a Redis string.
Con
Cannot get individual JSON properties

Cache strategies
Application objects
Key: CUSTOMER_ID:123,
Value: Customer object as byte array
String firstName = rs.getString(First_Name);
customer.setFirstName(firstName);
String lastName = rs.getString(Last_Name);
customer.setLastName(lastName);
Pro
Use application objects in their native structure and data state when serialized

Cache strategies
Using Redis data structures
Key: CUSTOMER_ID:123, Value: rsHash
String firstName = rs.getString(First_Name);
rsHash.put(“firstName", firstName);
String lastName = rs.getString(Last_Name);
rsHash.put(“lastName", lastName);
jedis.hmset(“CUSTOMER_ID:123", rsHash);
Pro
In addition to reducing data retrieval latency, cache data into specific data structure that simplifies the
data access pattern

• Understand the frequency of change of underlying data
• Set appropriate TTLs on keys that match that frequency
• Choose appropriate eviction policies that are aligned with application requirements
• Isolate your cluster by purpose (for example, cache cluster, queue, standalone database, and so on)
• Maintain cache freshness with write-throughs
• Performance test and size your cluster appropriately
• Monitor Cache HIT/MISS ratio and alarm on poor performance
• Use failover API to test application resiliency
Caching tips

Cluster sizing best practices
• Storage—clusters should have adequate memory
• Recommended: memory needed + 25% reserved memory (for Redis) + some room for growth
(optional 10%)
• Optimize using eviction policies and TTLs
• Scale up or out before reaching max-memory using CloudWatch alarms
• Use memory optimized nodes for cost effectiveness (R4 support)
• Performance—performance should not be compromised
• Benchmark operations using Redis Benchmark tool
• For more READIOPS—add replicas
• For more WRITEIOPS—add shards (scale out)
• For more network IO—use network optimized instances and scale out
• Use pipelining for bulk reads/writes
• Consider Big(O) time complexity for data structure commands
• Cluster isolation (apps sharing key space)—choose a strategy that works for your workload
• Identify what kind of isolation is needed based on the workload and environment
• Isolation: No Isolation $ | Isolation by Purpose $$ | Full Isolation $$$

Redis benchmark tool
Open source utility to benchmark performance
example: src/redis-benchmark -h r3-xlarge-perf.foio87.0001.use1.cache.amazonaws.com -p 6379 -n -150000 -d 100
Syntax:
redis-benchmark -h <host> -p <port> -c 50 -n 1000 -d 500 –q
-c <clients>—Specifies the number of parallel connections (default 50).
-n <requests>—Specifies the number of requests (default 1000000).
-d <size>—Specifies the data size of GET and SET values in bytes.
-t <test1,test2>—Comma-separated list of tests to perform.
-q—Quiet operation, displays only the result.

Redis max-memory policies
Select a max-memory policy based on your workload needs
• noeviction: return errors when the memory limit has been reached and the client is trying to execute
commands that might result in more memory to be used
• allkeys-lru: evict keys trying to remove the less recently used (LRU) keys first
• volatile-lru: evict keys trying to remove the less recently used (LRU) keys first, but only among keys
that have an expire set
• allkeys-random: evict random keys to make space for the new data added
• volatile-random: evict random keys to make space for the new data added, but only evict keys with an
expire set
• volatile-ttl: evict only keys with an expire set, and try to evict keys with a shorter time to live (TTL) first

Key ElastiCache CloudWatch metrics
• CPUUtilization
• Memcached—up to 90% ok
• Redis—divide by cores (ex: 90% / 4 = 22.5%)
• SwapUsage low
• CacheMisses/CacheHits Ratio low/stable
• Evictions near zero
• Exception: Russian-doll caching
• CurrConnections stable
• Setup alarms with CloudWatch metrics

ElastiCache modifiable parameters
• Maxclients: 65000 (unchangeable)
• Use connection pooling
• timeout—closes a connection after it has been idle for a given interval
• tcp-keepalive—detects dead peers given an interval
• Databases: 16 (default) for non-clustered mode
• Logical partition
• Reserved-memory: 25% (default)
• Recommended
§ 50% of maxmemory to use before 2.8.22
§ 25% after 2.8.22—ElastiCache
• Maxmemory-policy:
• The eviction policy for keys when maximum memory usage is reached
• Possible values: volatile-lru, allkeys-lru, volatile-random, allkeys-random, volatile-ttl,
noeviction

What we’re building
Availability Zone #1
REINVENT-RDS-SG
Test
Instance
VPC: REINVENT 10.0.0.0/16
Public1:
10.0.0.0/24
Private1: 10.0.2.0/24
REINVENT-EC-SG
REINVENT-APP-SG
APACHE HTTP SERVER
REDIS 3.2 CLIENT
JMETER 3.2
JUNIT 4 TEST APP
JAVA 1.8.0
Test
Instance
SSH,
HTTP
Performance testing lab—infrastructure
topology

Workshop: JMeter and JUnit tests overview
RDSLoad:testRDSLoad() (This executes only one time)
1. Drops database CustomerDB if it exists, referencing SQL stored in DropCustomerDB
2. Creates database CustomerDB, referencing SQL stored in CreateCustomerDB
3. Creates table customer, referencing SQL stored in CreateCustomerTbl
4. Loads 1000 customer records stored in CUSTOMER.sql
RedisLoad:testRedisLoad() (This executes only one time)
1. Flushes all the data in the Redis Cluster to prepare it
2. Queries the customer database table for the total count of customer records, referencing SQL stored in
CountCustomerTbl
3. Iterates over the total count and queries for each row, referencing SQL stored in SelectCustomerTbl
4. Stores each SQL ResultSet row into a Redis HashMap with the key name being the SQL statement for
that individual customer query
Data loading

Stress testing: 20k requests/second
RDSBenchmark:testRDS()
1. Randomly selects a customer row to retrieve from customer IDs 1–1000
2. Retrieves the SELECT SQL STATEMENT and appends the random customer ID integer referencing SQL stored in
SelectCustomerTbl
3. Executes SQL statement and iterates/displays the fetched row
RedisBenchmark:testRedis()
1. Randomly selects a customer hashmap to retrieve from customer IDs 1–1000
2. Retrieves the SELECT SQL STATEMENT and appends the random customer ID integer referencing SQL stored in
SelectCustomerTbl
3. Executes Redis HASHMAP HGETALL command to retrieve/display the fetched hash

//Generating a random num 1-1000 representing customer ids
randomKey = Integer.toString(rand.nextInt(1000));
//Retrieving Customer Select SQL Statement & appending customer id
sql = query + randomKey;
stmt = connection.createStatement();
//executing query
rs = stmt.executeQuery(sql);
//iterating and displaying results
while( rs.next() ) { …
//Generating a random num 1-1000 representing customer ids
randomKey = Integer.toString(rand.nextInt(1000));
//Retrieving Customer Select SQL Statement & appending customer id
key = query + randomKey;
//executing command
map = jedis.hgetAll(key);
//iterating and displaying results
for (String name : map.keySet() ) { …

Building High Performance Apps with In-memory Data

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

Similaire à Building High Performance Apps with In-memory Data

Similaire à Building High Performance Apps with In-memory Data (20)

Plus de Amazon Web Services

Plus de Amazon Web Services (20)

Building High Performance Apps with In-memory Data