Immutable data architecture with Datomic, Spark and Kafka

Konrad Scorciapino @konr

Mauro Lopes @maurolopes23

Share our experience

  • Data architecture
  • Supporting machine learning
  • Data access

Overview

0. Background
1. Scoring
2. Training
3. Analysing

Background

Microservices

An interaction

Database

  • Not SQL
  • Not NoSQL
  • Datomic

Board

Board

Piles

Pile

Pile

How is it stored?

[[1000 :person/name "Quercus"     1]
 [1000 :person/job  "Philosopher" 1]
 [1000 :person/city "São Paulo"   1]
 [1000 :person/city "Richmond"    2]]

How is it queried?

{:find [?job]
 :in  [?name]
 :where [[?p :person/name ?name]
         [?p :person/job ?job]]}

How is it queried?

{:find [?name]
 :in  [?job]
 :where [[?p :person/name ?name]
         [?p :person/job ?job]]}

How is it queried?

{:find [?name ?job]
 :in  []
 :where [[?p :person/name ?name]
         [?p :person/job ?job]]}

Scoring

(approach)

Model training vs Scoring

train(x_0, x_1, ..., x_m) -> Blackbox
train(x0,x1,...,xm)−>Blackboxtrain(x_0, x_1, ..., x_m) -> Blackbox
Blackbox(x) -> Score
Blackbox(x)−>ScoreBlackbox(x) -> Score

Scoring

Blackbox(x) -> Score
Blackbox(x)−>ScoreBlackbox(x) -> Score

How do we get   ?

x
xx

Entity from a queue

Enriched entity

Downsides

  • Affects production
  • Complexity on services
  • Traceability (Data, PII)

A second approach:

data directly from the database

Cursors

Cursor

Cursor

Entity from cursor and id

Multiple DBs

Multiple DBs

Scoring

(services)

A service where

  • All the data is avaliable
  • All services become one
  • All the chaos intertwines

MORDOR

One service to query them all

  • Read-only querying service
  • No interference

No interference

No interference

No interference

Using it

Sample message

Sample message

Sample query

{:as-of #inst "2014-11-15T13:37:42Z"
 :args ["pur-8"]
 :query
   {:find  [?name (max ?amount)]]
    :in    [?purchase-id]
    :where [[$pur ?purchase  :purchase/id ?purchase-id]
            [$pur ?purchase  :purchase/customer-id ?customer-id]
            
            [$cus ?customer  :customer/id ?customer-id]
            [$cus ?customer  :customer/name ?name]

            [$pur ?purchase2 :purchase/customer-id ?customer-id]
            [$pur ?purchase2 :purchase/amount ?amount]]}}

Model service

Output

Training

Goal

train(x_0, x_1, ..., x_m) -> Blackbox
train(x0,x1,...,xm)−>Blackboxtrain(x_0, x_1, ..., x_m) -> Blackbox
Blackbox(x) -> Score
Blackbox(x)−>ScoreBlackbox(x) -> Score

Scoring time

Training time

Training time

Apache Spark

  • Large scale data processing
  • High performance
  • Easy to use
  • Cluster

RDD

RDD

RDDs: our use case

RDDs: our use case

Sharding queries

  • Regular query
  • Base entity
  • Filter using mod

Sharding queries

{:query
   {:find  [?purchase-amount ?bill-amount]]
    :where [[$pur ?purchase :purchase/id]
            [$pur ?purchase :purchase/bill-id ?bill-id]
            [$pur ?purchase :purchase/amount ?purchase-amount]

            [$bil ?bill :bill/id ?bill-id]
            [$bil ?bill :bill/amount ?bill-amount]]}}

Sharding queries

{:query
   {:find  [?purchase-amount ?bill-amount]]
    :in    [?shard-index ?shard-count]                  ;; <--------
    :where [[$pur ?purchase :purchase/id]
            [(mod ?purchase ?shard-count) ?shard-index] ;; <--------
            [$pur ?purchase :purchase/bill-id ?bill-id]
            [$pur ?purchase :purchase/amount ?purchase-amount]

            [$bil ?bill :bill/id ?bill-id]
            [$bil ?bill :bill/amount ?bill-amount]]}
 :args [0 35]}                                          ;; <--------

Analysing

Data access

Data access

Learning curve

{:as-of #inst "2014-11-15T13:37:42Z"
 :args ["pur-8"]
 :query
   {:find  [?name (max ?amount)]]
    :in    [?purchase-id]
    :where [[$pur ?purchase  :purchase/id ?purchase-id]
            [$pur ?purchase  :purchase/customer-id ?customer-id]
            
            [$cus ?customer  :customer/id ?customer-id]
            [$cus ?customer  :customer/name ?name]

            [$pur ?purchase2 :purchase/customer-id ?customer-id]
            [$pur ?purchase2 :purchase/amount ?amount]]}}

Improving experience through saved queries

  • Cross-functional teams
  • Tech-savvy people create and save queries
  • Other people reuse and learn
  • Share data and procedures

UI

Usage statistics

  • Used by ALL teams
  • More than one million query executions
  • 100s (usually 1000s) of queries by each team
  • 618 saved queries

Stored procedures

  • random_numbers()
  • interest()
  • late()

Testimonials

  • "<Our tooling> helps me be more assertive and back hypotheses with data from queries I write myself (...) I can get a snapshot of the data and then follow its evolution over a period of time" - Business analyst

  • "<Our tooling> helps me find corner cases, trace back the origin of data, and figure out why it ended up that way (...)"  - Software Engineer

  • "I like how it's both a querying and analysing tool (...) having a timestamp from an analysis I made, I can choose to reproduce the results or redo it using up-to-date data just minutes before a meeting" - Data analyst

Final remarks

What we have done

  • Solution for scoring
  • Solution for training
  • Solution for data access

What we didn't have to

  • Copy data around
  • Duplicate functions/logic
  • Create views

:)

Made with Slides.com