Date post: | 12-Jul-2015 |
Category: |
Technology |
Upload: | juan-valencia |
View: | 1,769 times |
Download: | 1 times |
1
Real-World Cassandra at ShareThis
Use Cases, Data Modeling, and Hector
ShareThis + Our Customers: Keys to Unlocking Social
2
1. DEPLOY SOCIAL TOOLS ACROSS BRANDS (AND DEVICES)
2. TAKE YOUR SOCIAL INVENTORY TO MARKET
3. LEVERAGE SHARETHIS: FOR DIRECT SALES, RESEARCH AND UN-RESERVED INVENTORY
Largest Ecosystem For Sharing and EngagementEngagement Across The Web
3
120 SOCIAL CHANNELS120 SOCIAL CHANNELS
SHARETHIS ECOSYSTEMSHARETHIS ECOSYSTEM
211 MILLION PEOPLE(95.1% of the web)
2.4 MILLION PUBLISHERS
Source: ComScore U.S. January 2013; internal numbers, January 2013
Data Modeling and Why it Matters (Keep it even, Keep it slice-able)
5
Use Cases
A New Product: SnapSets
3 - x1.large
A New Product: SnapSets
Use Case: SnapSets, A New Product
Use Case: SnapSets, A New Product (Continued)
CF: Users (userId)meta:first_name=Ronaldmeta:last_name=Melenciometa:username=ronsharethisscrapbook:timestamp:scrapbookId:name=Scrapbook 1scrapbook:timestamp:scrapbookId:date_created=Jan 10url1:sid:clipID={LOCATION DATA}url1:sid:456={LOCATION DATA}
CF: Scrapbooks (scrapbookId)clip:timestamp:clipId:url=sharethis.comclip:timestamp:clipId:title=Clip 1clip:timestamp:clipId:likes=10
CF: Clip (clipId)comment:timestamp:commentId={"name":"Ronald","timestamp":'"jan 10","comment":"hi"}
CF: Stats (user:userId,application,publisher:pubId)meta:total_scrapbooks=1meta:total_clips=100meta:total_scrapbook_comments=100scrapbook:timestamp:scrapbookId:total_comments=10scrapbook:timestamp:scrapbookId:clip:timestamp:clipId:likes=10scrapbook:timestamp:scrapbookId:clip:timestamp:clipId:dislikes=10
9
Use Cases
High Velocity Reads and Writes: Count Service
9 – hi1.4xlarge9 – x1.large
Use Case: Count Service for URL's
● 1 Billion Pageviews per day = 12k pageviews per second
● 60 Million Social Referrals per day = 720 social referrals per second
● 1 Million Shares per day = 12 shares per second
● No expiration of Data* (3bn rows)
● Requires minimum latency possible
● Multiple read requests per page on blogs
● Normalize and Hash the URL for a row key
● Each social channel is a column
● Retrieve the whole row for counts
● Fix it by cheating ^_^ *
12
Use Cases
Insights that Matter – Your Social Analytics Dashboard
13
Timely Social Analytics
Dive deeper into your most social content
Identify popular articles
Uncover which social channels are driving
the most social traffic
Benchmark your social engagement with SQI
Measure social activity on an hourly, daily, weekly & monthly basis.
12 - x1.large
Use Case: Loading Processed Batch Data
● Backend Hadoop stack for processing analytics
● 58 JSON schemas map tabular data to key/value storage for slicing
● MondoDB* did not scale for frequent row level writes on the same table
● Needed to maintain read throughput during spikes to writes when analytics were finished
● No TTL* - Works daily, doesn't work hourly
● Switching from Astyanax to Hector
● Using a Hector Client through Java API's
Use Case: Loading Processed Batch Data (continued)
{ "schema": [ { "column_name":"publisher", "column_type":"UTF8Type", "column_level":"common", "column_master":"" }, {"column_name":"domain","column_type":"UTF8Type","column_level":"common","column_master":""}, {"column_name":"percenta","column_type":"FloatType","column_level":"composite_slave","column_master":"category"}, {"column_name":"percentb","column_type":"FloatType","column_level":"composite_slave","column_master":"category"}, {"column_name":"sqi","column_type":"FloatType","column_level":"composite_slave","column_master":"category"}, {"column_name":"month","column_type":"UTF8Type","column_level":"partition","column_master":""}, {"column_name":"category","column_type":"UTF8Type","column_level":"composite_master","column_master":""} ], "row_key_format": "publisher:domain:month", "column_family_name": "sqi_table"}
CF -> Data TypeRow -> Publisher:domain:timestampColumns -> master:slave = value (topics, categories, urls, timestamps, etc)
16
Use Cases
Insights that Matter – Your Social Analytics Dashboard
17
Real Time Social Analytics
Dive deeper into your most social content
Identify trending articles in real-time
Uncover which social channels are driving
the most social traffic
Benchmark your social engagement with SQI
Measure social activity on an hourly, daily, weekly & monthly basis.
12 - cc1.4xlarge
Insights that Matter – Your Social Analytics Dashboard
18
Real Time Social Analytics
Dive deeper into your most social content
Identify trending articles in real-time
Uncover which social channels are driving
the most social traffic
Benchmark your social engagement with SQI
Measure social activity on an hourly, daily, weekly & monthly basis.
12 - cc1.4xlarge
Insights that Matter – And aren't accessible
Insights that Matter – And aren't accessible
Insights that Matter – And aren't accessible
● Too many columns – unbounded url / channel sets
● Cascading failure
● Solutions:
– Bigger Boxes – meh...
– Split up the columns – split the rowkeys
● Hash Urls and keep stats separate
– Split up the columns – split the CF
● Move URLs to their own space
– Split up the columns – split the Keyspace
● Keyspace is a timestamp
22
Ask Good Data Modeling Questions
23
● How many rows will there be?● How many columns per row will you need?● How will you slice your data?● What are the maximum number of rows ?● What are the maximum number of columns?● Is your data relational?● How long will your data live?
24
Hectorhttps://github.com/hector-client/hector/wiki/User-Guide
Hector Imports
import me.prettyprint.cassandra.model.BasicColumnFamilyDefinition;import me.prettyprint.cassandra.model.ConfigurableConsistencyLevel;import me.prettyprint.cassandra.serializers.LongSerializer;import me.prettyprint.cassandra.serializers.StringSerializer;import me.prettyprint.cassandra.service.ColumnSliceIterator;import me.prettyprint.cassandra.service.ThriftCfDef;import me.prettyprint.cassandra.service.ThriftKsDef;import me.prettyprint.cassandra.service.template.ColumnFamilyResult;import me.prettyprint.cassandra.service.template.ColumnFamilyTemplate;import me.prettyprint.cassandra.service.template.ThriftColumnFamilyTemplate;
import me.prettyprint.hector.api.beans.ColumnSlice;import me.prettyprint.hector.api.beans.HColumn;import me.prettyprint.hector.api.beans.HCounterColumn;import me.prettyprint.hector.api.ddl.ColumnFamilyDefinition;import me.prettyprint.hector.api.ddl.ComparatorType;import me.prettyprint.hector.api.ddl.KeyspaceDefinition;import me.prettyprint.hector.api.exceptions.HectorException;import me.prettyprint.hector.api.factory.HFactory;import me.prettyprint.hector.api.mutation.Mutator;import me.prettyprint.hector.api.query.ColumnQuery;import me.prettyprint.hector.api.query.CounterQuery;import me.prettyprint.hector.api.query.QueryResult;import me.prettyprint.hector.api.query.SliceCounterQuery;import me.prettyprint.hector.api.query.SliceQuery;
Hector: Add a keyspace
public static Cluster getCluster(String name, String hosts) { return HFactory.getOrCreateCluster(name, hosts); }
public static KeyspaceDefinition createKeyspaceDefinition(String keyspaceName, int replication) { return HFactory.createKeyspaceDefinition( keyspaceName, ThriftKsDef.DEF_STRATEGY_CLASS, // "org.apache.cassandra.locator.SimpleStrategy" replication, null // ArrayList of CF definitions ); }
public static void addKeyspace(Cluster cluster, KeyspaceDefinition ksDef) { KeyspaceDefinition keyspaceDef = cluster.describeKeyspace(ksDef.getName()); if (keyspaceDef == null) { cluster.addKeyspace(ksDef, true); System.out.println("Created keyspace: " + ksDef.getName()); } else { System.err.println("Keyspace already exists"); } }
Hector: Define a CF
public static ColumnFamilyDefinition createGenericColumnFamilyDefinition( String ksName, String cfName, ComparatorType ctName) { BasicColumnFamilyDefinition columnFamilyDefinition = new BasicColumnFamilyDefinition(); columnFamilyDefinition.setKeyspaceName(ksName); columnFamilyDefinition.setName(cfName); columnFamilyDefinition.setDefaultValidationClass(ctName.getClassName()); columnFamilyDefinition.setReplicateOnWrite(true); return new ThriftCfDef(columnFamilyDefinition); } public static ColumnFamilyDefinition createCounterColumnFamilyDefinition(String ksName, String cfName) { BasicColumnFamilyDefinition columnFamilyDefinition = new BasicColumnFamilyDefinition(); columnFamilyDefinition.setKeyspaceName(ksName); columnFamilyDefinition.setName(cfName); columnFamilyDefinition.setDefaultValidationClass(ComparatorType.COUNTERTYPE.getClassName()); columnFamilyDefinition.setReplicateOnWrite(true); return new ThriftCfDef(columnFamilyDefinition); }
Hector: Add a CF
Keyspace k = HFactory.createKeyspace(nameString, cluster);
public static void addColumnFamily(Cluster cluster, Keyspace keyspace, ColumnFamilyDefinition cfDef) { KeyspaceDefinition ksDef = cluster.describeKeyspace(keyspace.getKeyspaceName()); if (ksDef != null) { List<ColumnFamilyDefinition> list = ksDef.getCfDefs(); String cfName = cfDef.getName(); boolean exists = false; for (ColumnFamilyDefinition myCfDef : list) { if (myCfDef.getName().equals(cfName)) { exists = true; System.err.println("Found Column Family: " + cfName + ". Did not insert."); } } if (!exists) { cluster.addColumnFamily(cfDef, true); System.out.println("Created column family: " + cfDef.getName()); } } else { System.err.println("Keyspace definition is null"); } }
Hector: Insert Column
public static void insertColumn( Cluster cluster, Keyspace keyspace, String cfName, String rowKey, String columnName, String columnValue) { Mutator<String> mutator = HFactory.createMutator(keyspace, StringSerializer.get()); //HFactory.createColumn(columnName, columnValue, StringSerializer.get(), StringSerializer.get()) HColumn<String, String> hCol = HFactory.createStringColumn(columnName, columnValue); mutator.insert(rowKey, cfName, hCol); mutator.execute(); }
public static void incrementCounter( Cluster cluster, Keyspace keyspace, String cfName, String rowKey, String counterColumnName) { Mutator<String> mutator = HFactory.createMutator(keyspace, StringSerializer.get()); mutator.insertCounter( rowKey, cfName, HFactory.createCounterColumn(counterColumnName, 1, StringSerializer.get())); mutator.execute(); }
Hector: Read Column
public static String getColumn( Cluster cluster, Keyspace keyspace,
String cfName, String rowKey, String columnName) { ColumnQuery<String, String, String> query = Hfactory.createColumnQuery(
keyspace, StringSerializer.get(), StringSerializer.get(), StringSerializer.get()); query.setColumnFamily(cfName).setKey(rowKey).setName(columnName); HColumn<String, String> value = query.execute().get(); if (value != null) { return value.getValue(); } return ""; }
Hector: Read Column
public static String getColumn( Cluster cluster, Keyspace keyspace,
String cfName, String rowKey, String columnName) { ColumnQuery<String, String, String> query = Hfactory.createColumnQuery(
keyspace, StringSerializer.get(), StringSerializer.get(), StringSerializer.get()); query.setColumnFamily(cfName).setKey(rowKey).setName(columnName); HColumn<String, String> value = query.execute().get(); if (value != null) { return value.getValue(); } return ""; }
Hector: Read Column
public static long getCounter( Cluster cluster, Keyspace keyspace,
String cfName, String rowKey, String counterColumnName) { CounterQuery<String, String> query =
HFactory.createCounterColumnQuery(keyspace, StringSerializer.get(),StringSerializer.get());
query.setColumnFamily(cfName).setKey(rowKey).setName(counterColumnName); HCounterColumn<String> counter = query.execute().get(); if (counter != null) { return counter.getValue(); } return 0; }
Hector: Read A Slice
public static Map<String, String> getSlice( Cluster cluster, Keyspace keyspace, String cfName, String rowKey, String start, String end, boolean reversed, int count) {
SliceQuery<String, String, String> query = HFactory.createSliceQuery(keyspace, StringSerializer.get(), StringSerializer.get(), StringSerializer.get());
// for counter use HFactory.createSliceQuery query.setColumnFamily(cfName); query.setKey(rowKey); query.setRange(start, end, reversed, count); Iterator<HColumn<String, String>> iter = query.execute().get().getColumns().iterator(); Map<String, String> answer = new HashMap<String, String>(); while (iter.hasNext()) { HColumn<String, String> temp = iter.next(); answer.put(temp.getName(), temp.getValue()); } return answer; }
Hector: Read All Columns
public static Map<String, String> getAllValues( Cluster cluster, String keyspace, String cf, String rowkey) {
HashMap<String, String> values = new HashMap<String, String>(); Keyspace keyspaceObject = HFactory.createKeyspace(keyspace, cluster); SliceQuery<String,String,String> query =
Hfactory.createSliceQuery(keyspaceObject, StringSerializer.get(), StringSerializer.get(), StringSerializer.get());
query.setColumnFamily(cf).setKey(rowkey).setRange("", "", true, 10000); QueryResult<ColumnSlice<String,String>> result = query.execute(); Iterator<HColumn<String, String>> iter = result.get().getColumns().iterator(); while (iter.hasNext()) { HColumn<String, String> current = iter.next(); values.put(current.getName(), current.getValue()); } return values; }
Hector: DANGER
private static void dropAllKeyspaces(Cluster cluster) { for (KeyspaceDefinition ksDef: cluster.describeKeyspaces()) { if (!(ksDef.getName().equals("system") || ksDef.getName().equals("OpsCenter"))) { cluster.dropKeyspace(ksDef.getName(), true); System.out.println("Dropped keyspace: " + ksDef.getName()); } } } private static void dropKeyspace(Cluster cluster, String keyspace) { KeyspaceDefinition ksDef = createKeyspaceDefinition(keyspace, Hector.replication); cluster.dropKeyspace(ksDef.getName(), true); System.out.println("Dropped keyspace: " + ksDef.getName()); } private static void dropColumnFamily(Cluster cluster, String keyspace, String cf) { cluster.dropColumnFamily(keyspace, cf); System.out.println("Dropped Column Family: " + cf ); }
Conclusions
● Data Modeling is Important
● Use Cassandra for write throughput
● Keep your ring even and your data slice-able
● Wrap your libraries and switch when you need to
● We're hiring: http://www.sharethis.com/about/careers
● Work with REAL big data, billions of requests per day
● Work on products that millions people see and interact with on a daily basis
● Work with a real-time pipeline, machine learning, complex user models
● #1 fastest growing company San Francisco
● free lunches
● ... and of course work with a bunch fun, smart people and PhDs
38
Thank You!