Using Databases to pull your application weight

This is the presentation that I took at south east ruby conference, October 5th 2017.

Transcript

1. Using Database to pull your application’s weight Harisankar P S, Founder & CEO, Red Panthers
2. { “name” => "Harisankar P S", “email” => ”mailme@hsps.in”, “twitter” => "coderhs", “facebook" => "coderhs", “github” => “coderhs”, “linkedin” => “coderhs”, }
3. One thing you notice about me is. I love Stickers!! So if you have stickers give them to me.
4. Ruby on Rails dev shop https://redpanthers.co
5. I am from Kochi, Kerala, India.
6. Interesting fact about India: India has 22 official languages,1653 spoken language and over 50,000 Dialects
7. So let me tell a story
8. about a novice developer
9. who wrote the first line of code
10. for a web app that was meant to process a
11. Couple of 1000 rows of data, less than 10 users
12. but grew so BIG that it was process GB’s of data every hour.
13. Me in 2014
14. This talk is about all the things that I learned Which helped me Scale the application without having to spend a fortune in Hardware
15. Let me start with mentioning the tools I used.
16. Really awesome when they work together
17. Ruby is Captain America
18. Rails is Iron man
19. Database is the hulk
20. Hulk can smash
21. But we make him carry our suitcase
22. This talk is about how we can offload couple of the jobs done by Rails to Database. You have a HULK, don’t feel scared to USE it.
23. I have done all these in production , so you don’t to feel scared to run this.
24. Today we are going to talk about • Query Planner • Indexing • Attribute Preloading • Materialised Views • Generating JSON • Synchronous Commit
25. Query Planner
26. Database is a General Purpose Software
27. A database is not build for a single use case or industry.
28. Then how does it handle all the scenarios?
29. Truth is, it doesn’t!
30. DB doesn’t know what all scenarios its put under, its upto us to nudge and optimise it.
31. • SQL syntax is all about how the results should be • What you want in your result - SELECT id, name • Or some information about data like - SELECT average(price), max(price), min(price) Where is the decision on how the data should be fetch made.
32. Well thats what Query planner is all about.
33. Its the Brain of your DB
34. We need to understand how the system work before we can improve its performance
35. • A query plan is created by the DB before the query you gave is executed. • Plan is the cost of running the query. The DB chooses the one with the least cost. • Query Plan assumes the plan it has is the ideal one
36. So we need to see what the query planner see Active Record has .explain method to help us there
37. Asset.where(asset_id: 1).explain User.where(id: 1).explain
38. So we check the query plan find where we are slowing down and then fix them and make the plan choose the faster method.
39. Sounds Simple Doesn't it
40. So lets see how we can do that.
41. Tip: We can make the query plan display in JSON, YML & XML formats as well EXPLAIN (format YAML) select * from users
42. Indexing
43. Okay..Lets not do that.
44. Indexes are a special lookup table that the database search engine can use to speed up data retrieval.
45. An Index is like a pointer to a particular row of a table. Where all the fields in the table are ordered.
46. But you know something?
47. Databases are smart
48. Even if you have indexes if it find the sequential search to be cost less then it would go for that one.
49. Example: Lets say you have a column with a 10,000 rows, but the the content is either short, medium, long. The database don’t use index as it finds sequential is faster
50. We should Index • Index Primary key • Index Foreign key • Index all columns you would be passing into where clause • Index the keys used to Join tables • Index the date column (if you are going to call it frequent, like rankings of a particular date) • Add partial index to scopes
51. Do not Index • Do not index tables with a lot of read, write • Do not index tables you know that will remain small, all through out its life time • Do not index columns where you will be manipulating lot of its values.
52. Attribute Preloading
53. A good use of Postgres Array
54. Rails Way
55. select * from tasks select * from tags inner join tasks_tags on tags.id = tasks_tags.tag_id where tasks_tags.task_id in (1,2,3,..) tasks = Task.find(:all, :include => :tags) 2 Queries Object for each tasks Rails Code
56. Fast Postgres Arrays: tasks = Task.find(:all, :select => "*, array(select tags.name from tags inner join tasks_tags on (tags.id = tasks_tags.tag_id) where tasks_tasks.task_id=tasks.id) as tag_names") 1 SQL query Rails doesn't have to create objects >3x faster
57. Materialised View
58. Database views? Database views are like the view in our rails. A rails view(an html page) shows data from multiple model in a single page Similarly we can show data from multiple table as a single table using the concept called views Why would we do that? Because it makes life easier
59. Instead of doing Every time you want the managers SELECT id, name, email FROM companies where role=‘manager’
60. CREATE VIEW company_managers AS SELECT id, name, email FROM companies WHERE role='manager'; You can create a view And simple do SELECT * FROM company_managers;
61. Note: • A schema of view lives in memory of a DB • The result is not stored in memory • Its is actually running our query to get the results • They are called pseudo tables
62. Materialised views are the next evolution of Database views. We store the result as well in a table • This was first introduced by Oracle • But now found in PostgreSQL, MicrosoftSQL, IBM DB2, etc. • MySQL doesn’t have it you can create it using open source extensions.
63. How can we use it in Ruby? Thanks to ActiveRecord its easy to access such pseudo tables
64. Create a migration to record the Materialised view We need a bit of SQL here class CreateAllTimesSalesMatView < ActiveRecord::Migration def up execute <<-SQL CREATE MATERIALIZED VIEW all_time_sales_mat_view AS SELECT sum(amount) as total_sale, DATE_TRUNC('day', invoice_adte) as date_of_sale FROM sales GROUP BY DATE_TRUNC('day', invoice_adte) SQL end def down execute("DROP MATERIALIZED VIEW IF EXISTS all_time_sales_view") end end
65. Create Active Record model I place these views at the location app/models/views class AllTimeSalesMatView < ActiveRecord::Base self.table_name = 'all_time_sales_mat_view' def readonly? true end def self.refresh ActiveRecord::Base.connection.execute('REFRESH MATERIALIZED VIEW CONCURRENTLY all_time_sales_mat_view') end end
66. Now we can do AllTimeSalesMatView.select(:name) AllTimeSalesMatView.where(email: 'hsps@redpanthers.co')
67. First, Last and Find • They don’t work in your view as they operate on your tables primary key and a view doesn’t have it • If you want to use it then you need to one of the fields in your table as primary key class Model < ActiveRecord::Base self.primary_key = :id end
68. Benchmark • I created a table with 1 million random sales and random dates in a year. (Dates where bookmarked as well)
69. Take Away • Faster to fetch data. • Capture commonly used joins & filters. • Push data intensive processing from Ruby to Database. • Allow fast and live filtering of complex associations or calculation .fields. • We can index various fields in the table.
70. Pain Points • We will be using more RAM and Storage • Requires Postgres 9.3 for MatView • Requires Postgres 9.4 to refresh concurrently • Can’t have Live data • You can fix this by creating your own table and 
 updating it with the latest information
71. JSON generation in DB
72. • Websites with simple HTML and plain javascript based AJAX is coming to an end • Its the era of new modern day JS frameworks • JSON is the glue that binds the fronted and our backend • So its natural to find more and more DB supporting the generation and storage of JSON.
73. To convert a single row to JSON select row_to_json(users) from users where id = 1 we use row_to_json() method in SQL
74. { “id":1,"email":"hsps@redpanthers.co", "encrypted_password":"iwillbecrazytodisplaythat", "reset_password_token":null,"reset_password_sent_at":null, "remember_created_at":"2016-11-06T08:39:47.983222", "sign_in_count": 11,"current_sign_in_at":"2016-11-18T11:47:01.946542", "last_sign_in_at":"2016-11-16T20:46:31.110257", "current_sign_in_ip":"::1","last_sign_in_ip":"::1", "created_at":"2016-11-06T08:38:46.193417", "updated_at":"2016-11-18T11:47:01.956152", “first_name":"Super","last_name":"Admin","role":3 }
75. But for more practical use we write queries like select row_to_json(results) from ( select id, email from users ) as results {"id":1,"email":"hsps@redpanthers.co"}
76. A more complex one select row_to_json(result) from ( select id, email, ( select array_to_json(array_agg(row_to_json(user_projects))) from ( select id, name from projects where user_id=users.id order by created_at asc ) user_projects ) as projects from users where id = 1 ) result
77. { “id":1,"email":"hsps@redpanthers.co", "project":["id": 3, "name": “CSnipp"] } We did data preloading as well, instead of having the need to run another query separate from the first one. We got the data about projects as well.
78. json_build_object • Added in PostgreSQL 9.4 to make JSON creation a bit more simpler select json_build_object('foo',1,'bar',2); {"foo": 1, "bar": 2}
79. So where is Ruby?
80. • For simple JSON creation you can use a gem called Surus • https://github.com/jackc/surus
81. Which lets you write code like User.find_json 1 User.find_json 1, columns: [:id, :name, :email] Post.find_json 1, include: :author User.find_json(user.id, include: {posts: {columns: [:id, :subject]}}) User.all_json User.where(admin: true).all_json User.all_json(columns: [:id, :name, :email], include: {posts: {columns: [:id, :subject]}}) Post.all_json(include: [:forum, :post])
82. But for more complicated queries you might still end up writing SQL
83. But Like me if you want to keep as much stuff as possible in Ruby then. Create a materialised view for your complicated query And then use the gem to generate JSON =)
84. Benchmarks • In our case we saw request to a (.json) url which used to take 2 seconds, coming down to <= 200ms • Some benchmarks I found online mentions
85. • Simple query • More complicate Query • Source: https://github.com/JackC/json_api_bench
86. Synchronous Commit
87. • PostgreSQL sacrifices speed for durability and reliability • PostgreSQL is known for its slow writes and faster readers • It has slow writes as it waits for confirmation that what we inserted has been recorded to the Hard Disk. • You can disable this confirmation check to speed up your inserts if you are inserting a lot of rows every second
88. User.transaction do User.synchronous_commit false @user.save end Surus Gem Provides
89. • Only issue now, is incase your DB crash it can’t recover the lost data not saved to Hard Disk • It won’t corrupt the data, but you might loose some rows of your data • Not to be used in cases when you want data integrity to be 100% • Use it where you don’t mind loosing some information or where you can rebuild it from outside your DB. Like logs, or raw information.
90. Postgres Config that we use # How much memory we have to cache the database, RAM_FOR_DATABASE * 3/4 effective_cache_size = <%= ram_for_database.to_i * 3/4 %>MB # Shared memory to hold data in RAM, RAM_FOR_DATABASE/4 shared_buffers = <%= ram_for_database.to_i / 3 %>MB # Work memory for queries (RAM_FOR_DATABASE/max_connections) ROUND DOWN 2^x work_mem = <%= 2**(Math.log(ram_for_database.to_i / expected_max_active_connections.to_i)/Math.log(2)).floor %>MB # Memory for vacuum, autovacuum, index creation, RAM/16 ROUND DOWN 2^x maintenance_work_mem = <%= 2**(Math.log(ram_for_database.to_i / 16)/Math.log(2)).floor %>MB # To ensure that we don't lose data, always fsync after commit synchronous_commit = on
91. # Size of WAL on disk, recommended setting: 16 checkpoint_segments = 16 # WAL memory buffer wal_buffers = 8MB # Ensure autovacuum is always turned on autovacuum = on # Set the number of concurrent disk I/O operations that PostgreSQL # expects can be executed simultaneously. effective_io_concurrency = 4
92. Summarize
93. • Index data so that we don’t end up scanning the whole DB • Use arrays for data preloading • Simplify the way you fetch data from the DB using views • Move complicated JSON generation to the Databases • Disable synchronous commit when you feel like it won’t cause a problem
94. Conclusions • Know your tech stack • We should have control over all our moving parts • Try to bring about the best with your tech stack before you start throwing more money at it • SQL has been around for 40 years and its planning to say for a while longer =) • There is no golden rule. What worked for me might not work for your specific use case.
95. I blogged about this in detail. • http://blog.redpanthers.co/materialized-views- caching-database-query/ • http://blog.redpanthers.co/create-json-response- using-postgresql-instead-rails/ • http://blog.redpanthers.co/different-types-index- postgresql/ • http://blog.redpanthers.co/optimising-postgresql- database-query-using-indexes/
96. Thank you ந"# നnി ਤuਹਾਡਾ ਧ'ਨਵਾਦ ಧನ#$ದ ధన ଥା# େଯାଉ ہیرکش ধন বাদ આભાર धन्यवाद अनुगृिहतोऽिस्म धन्यवाद ତu#$ୁ ଧନ(ବା+ ينابرهم يج ناهوت धन्यवाद * India has 1653 spoken languages