![Graph Model (1)
● Nodes and Relationships
○ (Machine)-[BELONGS_TO_CATEGORY]->(Category)
○ (Machine)-[IS_OF_MODEL]->(Model)
○ (Maker)-[MADE_MODEL]->(Model)
○ (Model)-[HAS_SPEC]->(Spec)
○ (Spec)-[BELONGS_TO_SPEC_CATEGORY]->(SpecCategory)](https://image.slidesharecdn.com/04-evolvingallstocker-191024090824/85/Evolving-ALLSTOCKER-Agile-increments-with-Pharo-Smalltalk-11-320.jpg)
![SCypher
user := 'user' asCypherObject.
friend := 'friend' asCypherObject.
friends := 'friends' asCypherObject.
query := CyQuery statements: {
CyMatch of: (CyRelationship start: user end: friend type: 'FRIEND').
CyWhere of: (CyExpression eq: (user prop: 'name') with: 'name'
asCypherParameter).
CyWith of: (user, ((CyFuncInvocation count: friend) as: friends)).
CyWhere of: (CyExpression gt: friends with: 10).
(CyReturn of: user) limit: 10.
}.
query cypherString. 'MATCH (user)-[:FRIEND]-(friend)
WHERE (user.name = $name)
WITH user, count(friend) AS friends
WHERE (friends > 10)
RETURN user LIMIT 10 '
● Generated query can be executed by
N4RestClient>>queryByCypher: queryString params: dictionary](https://image.slidesharecdn.com/04-evolvingallstocker-191024090824/85/Evolving-ALLSTOCKER-Agile-increments-with-Pharo-Smalltalk-15-320.jpg)
![RediStick
● A redis client supporting auto-reconnect
○ Reliable pubsub by pinging to itself
stick := RsRediStick targetUrl: 'sync://localhost'.
stick connect.
stick beSticky. "Auto reconnect when server is not accessible"
stick onError: [ :e | e pass ].
stick endpoint subscribe: #('ping') callback: [:msg | msg inspect].
"From another stick"
pubStick endpoint publish: 'ping' message: 'OK?'](https://image.slidesharecdn.com/04-evolvingallstocker-191024090824/85/Evolving-ALLSTOCKER-Agile-increments-with-Pharo-Smalltalk-24-320.jpg)
Evolving ALLSTOCKER: Agile increments with Pharo Smalltalk
- 1. Evolving ALLSTOCKER Masashi Umezawa & Kazunori Ueda ESUG 2019 Agile increments with Pharo Smalltalk
- 2. What is ALLSTOCKER? ● Online platform for trading used construction equipment and machinery ○ Marketplace ○ Real-time Bid Auction ● Over 4000 worldwide buyers ● Over 400 machines/month listed on the site ● Most systems are built with Pharo Smalltalk
- 3. Our Development Process - KANBAN Backlog Analysis Development Test Deploy ○ Swarm to make the flow smoother! →→→ →
- 4. 2015/02- In the Beginning ● First prototype was made in two weeks ○ Only 4 prerequisites (Seaside, Glorp, Nagare, AWS SDK for Smalltalk) ○ User/Machine registration, photo uploader, watchlist ● After 90 releases ○ 34 prerequisites ○ 1100+ classes
- 5. Our policy - Smalltalk as a Hub ● We take polyglot microservices approach ○ Programming languages ■ Smalltalk, JavaScript, Ruby, Lua, Groovy, Erlang, Python ○ Databases ■ PostgreSQL, Redis, Neo4j, Tarantool, MongoDB ○ External APIs ■ Elasticsearch, SendGrid, Mixpanel, Fluentd ● Smalltalk is a great hub for leveraging these elements
- 7. Why Search is Important? ● In our site, Search event ratio is 43+ percent ● Buyers first search, then view, and buy Search! View
- 8. 2015/7 - Full-Text Search ● Basic full-text search ● Elasticsearch-smalltalk for multilingual full-text search ○ Different analyzers for each language ○ Search results are boosted according to the user’s primary language
- 9. Elasticsearch-Smalltalk prepareSearch esSearch := ESSearch new index: self index. esSearch minScore: self minScore. ^esSearch search es := self prepareSearch. es query: self buildQuery. ^ es searchFrom: self offset size: self limit buildNameMatchQuery: words fieldName: fieldName ngramBoost: boostValue | phraseQuery matchQuery prefixQuery | phraseQuery := ESMatchQuery new matchPhrase; field: fieldName; query: words; yourself. prefixQuery := ESPrefixQuery new field: fieldName; query: words; yourself. matchQuery := ESMatchQuery new field: fieldName, '__ngram'; query: words; boost: boostValue; yourself. ^ ESBoolQuery new should: {phraseQuery. prefixQuery. matchQuery}; minimum_should_match: 1; yourself. ● Building ● Searching Boosted!
- 10. 2016/2 - Advanced Search ● Let’s support advanced search! ○ Many aggregation options ■ by category, maker, model number ● Elasticsearch was not enough to do complex aggregations ● Joining tables with Glorp was hard to maintain ● We adopted Graph database - Neo4j
- 11. Graph Model (1) ● Nodes and Relationships ○ (Machine)-[BELONGS_TO_CATEGORY]->(Category) ○ (Machine)-[IS_OF_MODEL]->(Model) ○ (Maker)-[MADE_MODEL]->(Model) ○ (Model)-[HAS_SPEC]->(Spec) ○ (Spec)-[BELONGS_TO_SPEC_CATEGORY]->(SpecCategory)
- 12. Neo4reSt ● Neo4j database client and Object wrappers ○ Introducing Neo4reSt db := N4GraphDb new. node1 := db createNode: {#name-> 'ume'}. node2 := db createNode: {#name-> 'Smalltalk'}. relation1 := node1 relateTo: node2 type: #uses properties: {#years->18}. db initialNode relateTo: node1 type: #people
- 13. Graph Model (2) ● We can freely get nodes/ relationships using Cypher query language
- 14. 2017/2 Revamping Advanced Search ● Need to generate complex queries dynamically according to various search options (spec filters) ● Hard-coded cypher queries were unmaintainable. ● SCypher was developed ○ “Manipulating Neo4j from Pharo Smalltalk” (Sample code project)
- 15. SCypher user := 'user' asCypherObject. friend := 'friend' asCypherObject. friends := 'friends' asCypherObject. query := CyQuery statements: { CyMatch of: (CyRelationship start: user end: friend type: 'FRIEND'). CyWhere of: (CyExpression eq: (user prop: 'name') with: 'name' asCypherParameter). CyWith of: (user, ((CyFuncInvocation count: friend) as: friends)). CyWhere of: (CyExpression gt: friends with: 10). (CyReturn of: user) limit: 10. }. query cypherString. 'MATCH (user)-[:FRIEND]-(friend) WHERE (user.name = $name) WITH user, count(friend) AS friends WHERE (friends > 10) RETURN user LIMIT 10 ' ● Generated query can be executed by N4RestClient>>queryByCypher: queryString params: dictionary
- 17. 2016/7-9- Beginning Realtime Auction System ● Need a highly reactive real-time bidding auction system ○ Vue.js (presentation) ○ Pharo (business logic) ○ Erlang (bidding core) ● Web API + Ajax + WebSocket
- 18. Original Auction System Network Architecture ● Bid requests ○ HTTP ● Notifications ○ WebSocket
- 19. 2018/3 Auction System Crisis ● Slow updates ● Too many connections ● Heavy load ● Difficult to log-in...
- 20. 2018/4-8 Scaling-out Auction System 1. On-demand notifications 2. Reducing the number of connections 3. Multiple Pharo images ● Our strategies:
- 21. On-demand Notifications ● Originally each client gets periodic updates on all items ○ Via Zinc-WebSocket ● Each client has observing item list ○ Client gets updates only when the item values are changed ○ Via Zinc-Server-Sent-Events
- 22. Reducing the Number of Connections ● SSE + HTTP/2 ○ Enables single TCP connection for many requests ● SSE is unidirectional and lightweight server { listen 443 ssl http2; ... }
- 23. Multiple Pharo images ● Let’s utilize multicore CPU! ● We divided one pharo image into three ○ auction-1, auction-2 (app server) ○ webhook (interact with Erlang bid server) ● Web API + Redis pubsub for mutual communication ○ RediStick pubsub mode
- 24. RediStick ● A redis client supporting auto-reconnect ○ Reliable pubsub by pinging to itself stick := RsRediStick targetUrl: 'sync://localhost'. stick connect. stick beSticky. "Auto reconnect when server is not accessible" stick onError: [ :e | e pass ]. stick endpoint subscribe: #('ping') callback: [:msg | msg inspect]. "From another stick" pubStick endpoint publish: 'ping' message: 'OK?'
- 25. Load-balancing via nginx ● Load-balancing by cookie-based sticky session upstream auction_upstream { hash $cookie_stocker consistent; server unix:/var/run/ auction_1.sock; server unix:/var/run/ auction_2.sock; } server { listen unix:/var/run/ auction_1.sock; location / { proxy_pass http:// as-auction-1.default.svc.cluster.local: 9000; } } server { … }
- 26. Revised Auction Architecture ●3 pharo images ●PUBSUB & SSE for async notifications
- 27. $ kubectl top pods NAME CPU(cores) MEMORY(bytes) as-auction-1-dpl-dc699554b-c5jh4 115m 261Mi as-auction-2-dpl-5d8bbd7b5c-qwlbp 111m 236Mi as-auction-webhook-dpl-66f47557b-jn2dc 48m 194Mi as-marketplace-dpl-7d898866d4-w72cc 119m 736Mi DONE!!
- 28. Questions? ● Visit allstocker.com ● Please stay tuned for more updates!