2. I’ll talk about
➔ Need for Tiered Data Stores in scaling systems and
Role of real time data change propagation systems.
◆ Example : Payments system
● Issues in old data tier architecture & motivation for
new one.
◆ Design Choices
➔ Aesop - Real Time Data Change Propagation System
◆ Aesop Scaling and High Availability
3. Payments - What does it do?
➔ Managing Transactions - Takes the customer through the life cycle of a Payment
Transaction.
➔ Reconciliation and Settlement - Reconcile with Bank and Settle to Merchants.
➔ Fraud Detection - Detect Payment Fraud
➔ Monitoring and Routing - Monitor for success rate of Transactions on various
dimensions and modifying routing.
4. Data Needs
Use Case Operation Requirement Data Retention
Transaction Flow Write + Read ACID + Normalised Structure
+ Low Latency
Transactions during Life Cycle +
All Data related to a Transaction
(Data for a Month)
Console Read + Search Denormalized Some attributes of a Transaction
Fraud Detection,
Financial Reports,
Monitoring
Aggregation + Unique
Values for a Transaction
Dimension
Aggregation and Large Data
access
1 year
Archival (Regulation) Reads, Reports Horizontally scalable data
store to store large amounts of
data.
All the Data
6. Payments - Old Data Tier Architecture
➔ MySql Master + Hot Standby + Slave.
➔ Application writes to Master. Transactional
and Real Time reads from Master.
➔ Historical Reads from Slave.
➔ Analytical and Aggregation queries onto
slave.
8. ➔ From these signs it was apparent that changes
were required in the data tier design.
➔ The current approach of one data store fits all
required change.
➔ The data tier would have to scale horizontally
and we needed more than one data store.
Multiple Data Stores
9. Multiple Data Stores - Issues?
➔ Data Consistency
➔ Real Time Data Availability across
stores
10. ETL?
Classic ETL approach has been around for decades and has a well defined and known solution. However this
was not an option for us because
➔ Data from the secondary stores is used to feed more than just business decisions.
➔ At Payments this data is supposed to feed into REAL TIME use cases like Console, Fraud Detection and
Monitoring Systems.
11. Dual Writes?
Application writes to destination data stores, synchronously or asynchronously. Application can write to a
Publisher-Subscriber system in which the Subscribers are consumers that eventually write to Destination Data
stores
➔ Pros : Appears Easy : Application can publish the same event that is being inserted/updated in the
Primary Data Source.
➔ Cons : Difficult to maintain consistency
◆ Writes are not Atomic - Ordering Issues
◆ Updates with non-primary-key where clause.
◆ Application Failures and Crashes.
◆ Manual changes in Primary Data Store will be missed.
12. Log Mining?
Log Mining
Separate application/service can extract changes from Database commit logs and publish them. This would use
the same approach used by database for replication.
➔ Pros : Consistency can be guaranteed as changes are being read from commit logs (bin log in case of
MySql).
➔ Cons
◆ Appears tough - But definitely possible.
◆ Tied to mechanism used by database for replication. Tied to commit log format, etc … Tightly
coupled approach.
Since Consistency across Datastores is of paramount importance to a financial system like Payments we chose
the Log Mining approach.
13. Approaches to Log Mining
MySql Bin Log Parsing
➔ Pros : Familiar approach
◆ Open source softwares were available
that parsed MySql bin logs. Open
Replicator and Tungsten Replicator
➔ Cons
◆ If format of bin logs changes the parser
would have to change.
◆ Open Replicator was supporting MySql
version 5.5. We would have to modify
Open Replicator to support MySql v5.6
and checksum feature introduced in
v5.6.
Custom Storage Engine
➔ Pros : Independent of binlog format. Layers
above Storage Engine take care of parsing.
➔ Cons : Unfamiliar approach. Unknown pitfalls.
Decided to go with known pitfalls and picked Bin Log
Parsing approach.
18. High Availability, Load Balancing and Scaling - Relay HA
Multiple Relay Servers read from the Source Data Sources.
➔ The Clients connect to Relay Server via a LB.
➔ Since the requests from clients are over HTTP one of the
Relay Servers or both can be serving the request based on
the configuration in the LB.
➔ When one Relay goes down the other can still handle the
requests.
19. Event Transformation
➔ Transforms the event as per the
mapping of source and destination
schema. It maps the source entity to
destination entity. The source
attribute is mapped to destination
attribute within the entity.
➔ A source entity can be mapped to
more than one destination entity
types.
➔ Map-All - one to one
➔ Hierarchical mapping
21. Summary
➔ Performance
◆ Relay : 1 XL VM (8 core, 32GB)
◆ Consumers : 4XL VM, 200 partitions
◆ Throughput : 20K-30K Inserts per sec
(MySQL to HBase)
◆ Data size : 500 GB
➔ What it is?
◆ Supports multiple data stores
◆ Delivers updates reliably - at least once
◆ Maintains Ordering within every
Partition
◆ Supports varying consumer speeds
➔ What is it not?
◆ Not exactly-once delivery
◆ Not a storage system
◆ No global ordering
➔ Support For
◆ Source
● MySql
● HBase
◆ Destination
● MySql
● HBase
● Elasticsearch
● Kafka
● Mapped Event Stream
22. More Details
➔ Project
◆ Open Source : https://github.com/Flipkart/aesop
◆ Support : aesop-users@googlegroups.com
◆ Multiple production deployments at Flipkart
➔ Related Work
◆ LinkedIn Databus
◆ Facebook Wormhole
➔ References
◆ Architecture of a Database System : http://db.cs.berkeley.edu/papers/fntdb07-architecture.pdf
◆ Wormhole Paper: https://www.usenix.org/system/files/conference/nsdi15/nsdi15-paper-sharma.pdf