No creation command ● You create an object by setting or adding to it ● Almost schema-less ● Can't use a command for one object to another
Redis keys all live in a single global namespace ● No schemas ● No separation by object type ● Very common pattern is to use fine grained keys, like (for a web session) web:111a7c9ff5afa0a7eb598b2c719c7975 ● KEYS command can find keys by pattern: ● KEYS web:* – Dangerous
How Redis users do “tables” ● They use a prefix: ● INCR hits:2013.05.25 ● They can find all these by doing ● KEYS hits:* ● Or they keep a set with all the keys for a given type of data ● SADD hitkeyset hits:2013.05.25 ● The application has to make use of these keys – Redis itself won't
Key prefix vs Key Set ● Key sets are much faster ● Ad server could not meet performance goals until it switched to using key sets ● Recommended by Redis docs
Using a key set to filter rows ● Sort of “where” clause ● Put the keys of the entries you want in a set somehow ● Can use command wrapper ● Define a new foreign table that uses that set as the keyset
9.3 notes ● In 9.3 there is json_populate_record() ● Could avoid use of hstore ● For post 9.3, would be a good idea to have a function converting an array of key value pairs to a record directly
Brand new – Singleton Key tables ● Each object is a table, not a row ● Sets and lists come back as single field rows ● Ordered sets come back as one or two field rows – second field can be score ● Hashes come back as rows of key/value
redis_connect() ● First argument is “handle” ● Remaining arguments are all optional ● con_host text DEFAULT '127.0.0.1'::text ● con_port integer DEFAULT 6379 ● con_pass text DEFAULT ' ::text ● con_db integer DEFAULT 0 ● ignore_duplicate boolean DEFAULT false
Redis wrapper connections are persistent ● Unlike FDW package, where they are made at the beginning of each table fetch ● Makes micro operations faster
redis_command and redis_command_argv ● Thin layers over similarly named functions in client library ● redis_command has max 4 arguments after command string – for more use redis_command_argv ● Might switch from VARIADIC text to VARIADIC “any”
Uses ● Push data into redis ● Redis utility statements from within Postgres
Higher level functions ● redis_push_record ● con_num integer ● data record ● push_keys boolean ● key_set text ● key_prefix text ● key_fields text
Why use Redis? ● Did I mention it's FAST? ● But not safe
Our use case ● An ad server for the web ● If Redis crashes, not a tragedy ● If it's slow, it's a tragedy
System Goals ● Serve 10,000 ads per second per application server cpu ● Use older existing hardware ● 5 ms for Postgres database to filter from 100k+ total ads to ~ 30 that can fit a page and meet business criteria ● 5 ms to filter to 1-5 best ads per page using statistics from Redis for freshness, revenue maximization etc. ● Record ad requests, confirmations and clicks. ● 24x7 operation with automatic fail over
Redundancy View Postgres 9.2 TransactionDB Cisco HSRP Node Pgpool Shorewal NGINX Hot Replication www.draw-shapes.de Business DB Multiple Hot Replication www.draw-shapes..de Instances Redis Keepalived Keepalived Skytools Londiste3 Data Warehouse DB Se S n e t n ine n l e Hot Replication Tier 1 Client Tier 2 Web Tier 3 Tier 4 Application Database
Postgres databases ● 6 Postgres databases ● Two for business model – master and streaming hot standby (small VM) ● Two for serving ads – master and streaming hot standby (physical Dell 2950) ● Two for for storing clicks and impressions – master and hot standby (physical Del 2950) ● Fronted by redundant pg pool load balancers with fail over and automated db fail over.
Business DB ● 30+ tables ● Example tables: ads, advertisers, publishers, ip locations ● Small number of users that manipulate the data (< 100) ● Typical application and screens ● Joining too slow to serve ads ● Tables get materialized into 2 tables in the ad serving database
Ad Serving Database ● Two tables ● First has ip ranges so we know where the user is coming from. Ad serving is often by country, region etc. ● Second has ad sizes, ad types, campaigns, keywords, channels, advertisers etc. ● Postgres inet type and index was a must have to be successful for table one ● Tsquery/tsvector, boxes, arrays were all a must have for table two (with associated index types)
Ad serving Database ● Materialized and copied from Business database every 3 minutes ● Indexes are created and new tables are vacuum analyzed then renamed. ● Performance goals were met. ● We doubt this could be done without Postgres data types and associated indexes ● Thanks
Recording Ad requests/confirmations and clicks ● At 10k/sec/cpu recording ads one row at a time + updates on confirmation is too slow ● Approach: record in Redis, update in Redis and once every six minutes we batch load from Redis to Postgres. - FDW was critical. ● Partitioning (inheritance) with constraint exclusion to segregate data by day using nightly batch job. One big table with a month's worth of data would not work. ● Table partitioning is not cheap in the leading commercial product.
Recording DB continued. ● Used heavily for reporting. ● Statistics tables (number of clicks, impressions etc.) are calculated every few minutes on today's data ● Calculated nightly for the whole day tables ● For reporting we needed some business data so we selectively replicate business tables in the ad recording database using Skytools. DB linking tables is too slow when joining.
Recording DB cont'd ● Another usage is fraud detection. ● Medium and long term frequency fraud detection is one type of fraud that this database is used for.
Redis ● In memory Database. ● Rich type support. ● Multiple copies and replication. ● Real time and short term fraud detection ● Dynamic pricing ● Statistical best Ad decision making ● Initial place to record and batch to Postgres ● Runs on VM with 94Gb of dedicated RAM.
Redis cont'd ● FDW and commands reduce the amount of code we had to write dramatically ● FDW good performance characteristics. ● Key success factor: In memory redis DB + postgres relational DB.
Postgres – Redis interaction ● Pricing data is pushed to Redis from Business DB via command wrapper ● Impression and Click data is pulled from Redis into Recording DB via Redis FDW
Current Status ● In production with 4 significant customers since March 1 ● Scaling wel
Conclusions ● Postgres' rich data types and associated indexes were absolutely essential ● Redis + Postgres with good FDW integration was the second key success factor ● Node.js concurrency was essential in getting good application throughput ● Open source allowed the system to be built for less than 2% of the cost of a competing commercial system