NoSQL: Now and Path Ahead
Shubham Kumar Srivastava
MakeMyTrip
Who am I
3
Abstract
What and Why : NoSql
Fundamentals
Use Case
Challenges
Path Ahead
.
What is NoSql
Database which does not adhere to the traditional relational database management system (RDMS) structure .
Why NoSql
Scalability and Performance
Cost
Data Modeling
Why NoSql : Motives and Drivers
Scalability and Performance
Horizontal scalability better than Vertical
Hardware getting cheaper and processing power increasing
Less Operational complexity as against RDBMS solutions.
In most of the solutions you get automatic sharding etc as default .
Why NoSql : Motives and Drivers contd..
Why NoSql : Motives and Drivers contd..
Why NoSql : Motives and Drivers contd.. Cost
Scale(as with NoSql) with Hefty Cost
Commodity hardware, software versions, upgrades, maintenance.
This brought organizations look out for alternatives and the need for a cost effective scale out option.
Why NoSql : Motives and Drivers contd.. Data ModelingSQL has been for
Concurreny,Consistency,Integrity
For Summations,Aggregations,Grouping’s
Schema Says: What all Do I answer ??
Why NoSql : Motives and Drivers contd.. Data Modeling
A plain key-value store is very powerful and fit the max use cases for a NoSQL solution
Hierarchical or graph-like data modelling and processing.
Values like maps of maps of maps.
Document Databases which even store arbitrary complex objects.
Document based indexing data store’s are a huge success.
Why NoSql : Motives and Drivers contd..At times SW apps are not limited to these constraints . This lead to data models like
Key/Value Store :
Redis,MemcacheDb/Voldemort etc.
Wide Column Store / Column Families : Cassandra/Hadoop(Hbase)/Hypertable/Cloudera etc.
Document Based Store’s :
Solr/Lucene/MongoDb/CouchDb/TerraStore etc.
Graph Data Store :
Neo4J/GraphBase/FlockDb etc.
Why NoSql : Motives and Drivers contd..
Schema Says: What are the questions
Data modeling is based on the set of Queries
Exploit De-normalization Duplication
Use Aggregates
Manage Joins with App + Aggregation + DeNormalization etc.
Why NoSql : Motives and Drivers contd..
Some Fanda-mentals
CAP Theorem
At the most only two properties of the three in a shared/distributed system can be satisfied.
Consistency
Availability
Tolerance to Network Partitions
CAP : Pictorially
Explanation
Use case: Scaling Web Apps
Critical fact’s : • Network outages are common • Customer shopping carts, email search, social network
queries—can tolerate stale data
How: Compromise on Consistency in-order to remain available vs disrupt user service at outages.
Rather than requiring consistency after every transaction, it is enough for the database to eventually be in a consistent state.
Brewer’s CAP theorem says you have no choice if you want to scale up.
Explanation
Explanation contd..
Sharp Contrast : High Speed Financial Application
Highly Transactional
Consistent
Automated
Can’t live with Eventual consistency
ACID vs BASE ACIDAtomic: Everything in a transaction succeeds or
the entire transaction is rolled back.
Consistent: A transaction cannot leave the database in an inconsistent state.
Isolated: Transactions cannot interfere with each other.
Durable: Completed transactions persist, even when servers restart etc.
Some Fanda-mentals cont..
BASEBasic Availability
Soft-state
Eventual consistency
Consistent Hashing
Common way to load balance .
The machine chosen to cache object o will be:
hash(o) mod n n:total number of machines
Consistent Hashing contd..
Adding a machine to the cache means hash(o) mod (n + 1)
Removing a machine to the cache means
hash(o) mod (n - 1)
Result on any above: Disaster
Swamped machines with redistribution
Consistent Hashing contd..
Commonly, a hash function(e.g MD5 hash) will
map a value into a 128-bit key, 0~2^127-1(or 32 bit even as given next) .
Consistent Hashing contd..
Consistent Hashing contd.. Both Key and Machine hashed with the same function
Consistent Hashing contd.. Adding a Node
Consistent Hashing contd.. Removing a Node
Use Case and NoSQL Solution
Problem:
Need to store bookings per day of all hotels . Queries centered around city and regions.
Hotel count : 1 Million
Date Range : Now to next 365 *2 Days
NoSQL: Path Ahead
ACID equivalence(Neo4J,CouchDb etc)
Transaction Support
Atomicity
MVCC
NoSQL: Path Ahead contd..
Possible Solution
Work with SQL Db w.r.t Creation/Updation etc.
Archive the data in NoSQL for query/analysis etc.
NoSQL: Path Ahead contd..
Enterprise Adoption and Challenges
NoSQL looks good for Unstructured data largely
SQL is the best choice for a broad range of traditional workloads.
NoSQL: Path Ahead contd..
NoSQL: Path Ahead contd..
Shout out loud Hybrid ACID + BASE They are not alternatives but
supplements
NoSQL: Path Ahead contd.. Maturity
Support
Skillset and Administration/Operation
Analytics and BI support
NoSQL: Path Ahead contd..
Q & A
References Nancy Lynch and Seth Gilbert, “Brewer's conjecture and the feasibility of consistent,
available, partition-tolerant web services”, ACM SIGACT News, Volume 33 Issue 2 (2002), pg. 51-59.
Brewer's CAP Theorem", julianbrowne.com, Retrieved 02-Mar-2010
Brewers CAP theorem on distributed systems", royans.net
CAP Twelve Years Later: How the "Rules" Have Changed on-line resource
E. Brewer, "Towards Robust Distributed Systems," Proc. 19th Ann. ACM Symp.Principles of Distributed Computing (PODC 00), ACM, 2000, pp. 7-10; on-line resource
D. Abadi, "Problems with CAP, and Yahoo’s Little Known NoSQL System," DBMS Musings, blog, 23 Apr. 2010; on-line resource.
C. Hale, "You Can’t Sacrifice Partition Tolerance," 7 Oct. 2010; on-line resource.
Facebook: Scaling Out on-line resource.
Gemstone : The Hardest Problems In Data Management on-line resource
The Log-Structured Merge-Tree (Research Paper)
CodeProject : Consistent Hashing on-line resource
HighlyScalable : NoSQL Data Modeling Techniques on-line resource
eBay Tech Blog :Cassandra Data Modeling Best Practices on-line resource
John D Cook : Acid Vs Base on-line resource
Merkle Trees
Phy-Accural Faliover Detaection (Research Paper)
Backup Slides
Better than the Original 1
Document Based DataStore{
_id : ObjectId("4e77bb3b8a3e000000004f7a"),
when : Date("2011-09-19T02:10:11.3Z",
author : "alex",
title : "No Free Lunch",
text : "This is the text of the post. It could be very long.",
tags : [ "business", "ramblings" ],
votes : 5,
voters : [ "jane", "joe", "spencer", "phyllis", "li" ],
comments : [
{ who : "jane", when : Date("2011-09-19T04:00:10.112Z"),
comment : "I agree." },
{ who : "meghan", when : Date("2011-09-20T14:36:06.958Z"),
comment : "You must be joking. etc etc ..." }
]
}
User and Items
User and Items : Option 1
User and Items : Option 2
User and Items : Option 3
User and Items : Option 4
Cassandra CF
Cassandra SuperCF
Use Case 1Ecommerce Site
Problem : Record User Preferences e.g : Location,IP,Currency selected, Source of Traffic, Multiple other dynamic values
Solution : In a CF based structure keep it simple
UserId_Key: Pref2_Name:Value1,Pref2_Name:Value2,….PrefN_Name:ValueN
Use Case 1RowKey: 1350136093705_6501082438199894
=> (column=1350136093764, value=-3242432#911167901131523, timestamp=1350136093766000)
=> (column=1350283322499, value=GOI#200701231712126570, timestamp=1350283322502001)
=> (column=1350283566051, value=GOI#200703221605283033, timestamp=1350283566054001)
=> (column=1350749595676, value=GOI#200805261514037199, timestamp=1350749595677001)
(column=1350785230322, value=BOM#200701251747233158, timestamp=1350785230324001)
RowKey: 1354499614310_10861558002828044
=> (column=1354499614368, value=TRV#201104071059204768, timestamp=1354499614370000, ttl=1728000)
-------------------
RowKey: 1349760150553_6114662943774777
=> (column=1349760152066, value=BLR#200802111324575807, timestamp=1349760152068001)
-------------------
RowKey: 1349805109805_6167423558533191
=> (column=1349805111833, value=TRV#312254274337517, timestamp=1349805111835001)
-------------------
RowKey: 1354435656227_7908056941568359
=> (column=1354435656367, value=IDR#200701211254519381, timestamp=1354435656369000, ttl=1728000)
-------------------
RowKey: 1347648097261_15570089270962881
=> (column=1347648097304, value=DEL#201101192008115545, timestamp=1347648097307000)
Use Case 1 Get
private Map<String, String> getPrerences(Keyspace keySpace, String userId, String... prefernceNames) throws IOException, CharacterCodingException {
SliceQuery<String, String, String> rsq = HFactory.createSliceQuery(keySpace, StringSerializer.get(), StringSerializer.get(), StringSerializer.get());
rsq.setColumnFamily(USER_PREFERENCE);
rsq.setKey(userId);
rsq.setColumnNames(prefernceNames);
QueryResult<ColumnSlice<String, String>> orows = rsq.execute();
Map<String, String> preferenceMap = new LinkedHashMap<String, String>();
for (HColumn<String, String> column : orows.get().getColumns()) {
preferenceMap.put(column.getName(), column.getValue());
}
return preferenceMap;
}
Use Case 1 Save
Mutator<String> m = HFactory.createMutator(keySpace, StringSerializer.get());
HColumn<String, String> userPrefrences = HFactory.createColumn(colkey, colvalue, StringSerializer.get(), StringSerializer.get());
userPrefrences.setTtl(ttlUserPrefrences);
m.addInsertion(rowkey, USER_PREFERENCE, userPrefrences);
m.execute();
Use Case 2
Online Travel Site
Problem: Need to know different metrics for a city hotels e.g.:
Hotels booked in last X TimeHotels Last viewed in Y TimeHotels Left with Z Inventory
Use Case 2RowKey: 2d323436353731
=> (super_column=911167901297486,
(column=6c6173747669657765646d657373616765, value=VIEWED#Last viewed 23 hour(s) ago., timestamp=1354962852610000)
column=6c6173747669657765646d657373616762, value=Inventory#20 , timestamp=1354962852610000,
column=6c6173747669657765646d657373616769, value=Bookings#8 , timestamp=135496282610000
)
-------------------
RowKey: 58524f
=> (super_column=200903041759196196,
(column=6c617374626f6f6b65646d657373616765, value=Booked#Last booked 1 day(s) ago., timestamp=1347781187842000)
(column=6c6173747669657765646d657373616765, value=VIEWED#Last viewed 2 hours ago., timestamp=1347707080147000))
=> (super_column=200903041848352230,
(column=6c6173747669657765646d657373616765, value=VIEWED#Last viewed 1 day(s) ago., timestamp=1347266107708000))
Use Case 2SuperSliceQuery<String, String, String, String> superQuery = HFactory.createSuperSliceQuery(getKeySpace(),
StringSerializer.get(), StringSerializer.get(),
StringSerializer.get(), StringSerializer.get());
superQuery.setColumnFamily(SUPER_SOCIAL_MESSAGE).setKey(cityCode);
QueryResult<SuperSlice<String, String, String>> result = superQuery.execute();
List<HSuperColumn<String, String, String>> superColumns = result.get().getSuperColumns();
if (superColumns != null) {
for (HSuperColumn<String, String, String> superColumn : superColumns) {
Map<String, String> messages = new HashMap<String, String>();
List<HColumn<String, String>> columns = superColumn.getColumns();
if (columns != null) {
for (HColumn<String, String> column : columns) {
messages.put(column.getName(), column.getValue());
}
}
/* The equivalent doc *\
document.addField(superColumn.getName(), messages);
documents.add(document);
}
}
Pig Script : MR <document>
<pigscript start="-16" end="-43200" start1="-1441" end1="-10080" start2="0" end2="-15" start3="0" end3="-1440">
<comment>Delete All Messages</comment>
<query><![CDATA[rows0 = LOAD 'cassandra://LH/HotelMessage' USING com.mmt.solr.hotels.cassandra.CassandraStorage() as (key:chararray, cols:bag{T:tuple(name:chararray, value:chararray) } );]]></query>
<query><![CDATA[cols0 = FOREACH rows0 GENERATE key as key,flatten($1) as (name:chararray, value:chararray);]]></query>
<query><![CDATA[cols0 = FOREACH rows0 GENERATE key as key,flatten($1) as (name:chararray, value:chararray);]]></query>
<query><![CDATA[userhotel0 = FOREACH cols0 GENERATE key as key,com.mmt.solr.hotels.cassandra.ByteBufferToString($1) as name,com.mmt.solr.hotels.cassandra.ByteBufferToString($2) as value;]]></query>
<query><![CDATA[uriCounts0 = FOREACH userhotel0 GENERATE key as citycode,com.mmt.solr.hotels.cassandra.ToBag(TOTUPLE(name,null));]]></query>
<comment>Last Viewed start 15 minutes to 30 days ago</comment>
<query><![CDATA[rows = LOAD 'cassandra://LH/LastViewedHotels?slice_start=#start&slice_end=#end&limit=1024&reversed=true' USING com.mmt.solr.hotels.cassandra.CassandraStorage() as (key:chararray, cols:bag{T:tuple(name:long, value:chararray) } );]]></query>
<query><![CDATA[cols = FOREACH rows GENERATE key as key,flatten($1) as (name:long, value:chararray);]]></query>
<query><![CDATA[userhotel = FOREACH cols GENERATE key as key,com.mmt.solr.hotels.cassandra.LongToHours($1) as name,com.mmt.solr.hotels.cassandra.ByteBufferToString($2) as value;]]></query>
<query><![CDATA[userhotelByCity = FOREACH userhotel GENERATE key as key,flatten($1) as name,flatten(org.apache.pig.piggybank.evaluation.string.Split(value,'#',2)) as (citycode:chararray,hotelid:chararray);]]></query>
<query><![CDATA[groupByhotels = GROUP userhotelByCity BY hotelid;]]></query>
<query><![CDATA[uriCounts = FOREACH groupByhotels { D = LIMIT userhotelByCity 1;
GENERATE flatten(D.citycode) as citycode,com.mmt.solr.hotels.cassandra.ToBag(
TOTUPLE(group,com.mmt.solr.hotels.cassandra.StringAppend('VIEWED#Last viewed ',D.name,' ago.')));
};]]></query>
<comment>Last Booked 1 to 8 days ago</comment>
<query><![CDATA[rows1 = LOAD 'cassandra://LH/BookedHotels?slice_start=#startA&slice_end=#endA&limit=1024&reversed=true' USING com.mmt.solr.hotels.cassandra.CassandraStorage() as (key:chararray, cols:bag{T:tuple(name:long, value:chararray) } );]]></query>
<query><![CDATA[cols1 = FOREACH rows1 GENERATE key as key,flatten($1) as (name:long, value:chararray);]]></query>
<query><![CDATA[userhotel1 = FOREACH cols1 GENERATE key as key,com.mmt.solr.hotels.cassandra.LongToHours($1) as name,com.mmt.solr.hotels.cassandra.ByteBufferToString($2) as value;]]></query>
<query><![CDATA[userhotelByCity1 = FOREACH userhotel1 GENERATE key as key,flatten($1) as name,flatten(org.apache.pig.piggybank.evaluation.string.Split(value,'#',2)) as (citycode:chararray,hotelid:chararray);]]></query>
<query><![CDATA[groupByhotels1 = GROUP userhotelByCity1 BY hotelid;]]></query>
<query><![CDATA[uriCounts1 = FOREACH groupByhotels1 { D = LIMIT userhotelByCity1 1;
GENERATE flatten(D.citycode) as citycode,com.mmt.solr.hotels.cassandra.ToBag(
TOTUPLE(group,com.mmt.solr.hotels.cassandra.StringAppend('Booked#Last booked ',D.name,' ago.')));
};]]></query>
Criteria's to Evaluate NoSQL Solutions
Internal partitioning
Automated flexible data distribution
Hot swappable nodes
Replication-style
Automated failover strategy