You've got a "situation"

• Data in enterprise systems
• Websites take a lot of effort
• Data access is slow
• Data distributed over several systems

Indexing Architecture

Workflow

Loaders

• Per external data source
• Lots of formats: SOAP, JSON, csv and XML files, Doctrine (Oracle)
• Modes
  • Push notifications
  • Changed since last successful run
  • Full import
  • Sanity checks: Compare item id list

Loaders

• Cronjobs hourly / daily / weekly
• Message queue workers for push
• All persist to MySQL and send message

Indexer

• Per API entity, aggregate and combine sources
• Mostly message queues
• Cronjobs for synchronising
• Sanity checks: Don't delete more than 10% of a type

Data Mapper

• Determine values for API models
• Modular, separate mapper per related fields
  • Testable code
  • Readable code
  • Reusable code
• Dependency modelling for partial updates

Queuing

• Record import state of each item
• When already waiting in queue, don't publish second message
• Track item state and processing delay

Retry queues

• Everything will go wrong once in a while
• On error, send message into retry queue
• Delayed for 4h, hope problem resolved by then
• Count retry and abort after 3 attempts
• Transparent from message consumer point of view

Notes on Elasticsearch

Indexes

• one index per language
• => one index per type
• de-normalize everything for fast responses
• no joins or parent-child relations

Schema Changes

• ES guesses field configurations
• Not always correctly => manual definitions
• No way to change definition of existing index
• Deploy new code with new index, but not yet online
• Copy from old index
  (or rebuild from MySQL)

Thank you!

@dbu

David Buchmann, Liip AG