Youtube Gaurav Sen

Messenging tools

Interview tips

Front to back approach: what user want -> how to configure servers
Each time you have a collection, think you should get a max number of items, size

Think about the features you want, 4 feature is enought for 1h interview
Store profile 1. Id 2. Score (good client ?) 3. Surname 4. Images (Give a max bound), Avatar
Recommend matches (Tinder)
Notes maches, analytices
Direct messaging (chat)
Group messages (Wts)
Sent / Delivered / Read notification (Wts)

Preparing beforehand at non peek hours
1. Identificate the hours
2. Cron jobs

5 tips video

Do not detail early (top-bottom) => speak alone
Don’t get a full architecture at once, wait for the scenario to evolve. Avoid silver bullet, golden harmer
Keep It Simple Stupid. Every block should be the same size
Justify each choice. Form your thought
Be tech aware (Name load balancer, database host)

Be 1. Clear 2. Flexible 3. Knowledge

Voc

Vertical scaling: buy a bigger machine
Horizontal scaling: buy more machines
Scalable: With horizontal, increase number of user
Consistent: the cache is consistent, the data if one server fails
Resilient: power failure, os crash
Fault tolerant
Resilient to changes
Extendable
Single point of failure: if one system fails, cannot be replaced
Microservice: A specialized program (cook empanadas) so each time a client ask for empanada, he can directly ask there
Decoupling, separate responsabilities: tje cook does no need to know if it is the.client a a delivery agent coming to get the empanada
Consisten hashing, used by load balancers to responsibilise a server on a (discontinous) range of request IDs
- Ensure there is no duplicayes
Notifier: Check heartbeat for each server (each 10sec) and if one is dead, notify
Database Sharding: Divide the data in different machines according to the mist relevant index => first find the most relvant index (goelic, age, tag). Also known as horizontal partitioning.
Client Server Communication Protocol: (ex: HTTP) The client requedt, the server respond. Pb: if slow operation, the client must request every 5s, generate noise
Peer to Peer: (ex: XMPP) Every actor is equal, WebSocket (C TCP)
Long pooling: repitively send HTTP request for user to check if there is a change in server (i.e. if he received a message)
Service discovery: each service.must authentificate (name, ip, port) to the load balancer
Health check, a 2 way heart beat to ensure a service is performing well his job, and kill it otherwise. Availability, economy of forces.

API

www.site.com/chat-messeging/getAdmins/v1 get m with gid=123 json in the request you can replace the get in the route because the HTTP request is get

For big response: 1. Pagination gives the responsibility to the client of repetitive calls 2. Fragmentation

Databases

Blob: Binary large object stored in databaese have 3 avantages over files
You must shard your database according to the query. Some frameworks do it automamatically (Casandra, Dynamo)
Blob
1. Mutability: can change cells := item attributes
2. Transaction safety: atomic operation safe
3. Indexablility
4. Access control: possible with files too
Files (No select *)
1. Cheap
2. Faster
3. Content Delivery Network: Distributed File System : each file is referenced by an url
Cache:
1. Save networkk call
2. Avoid repeated computation
3. Reduce Db load
Cache policies:
1. Least Recently Used (queue)
2. Least Frecuently Used
Redis is a global cache, slower but more accurate than in memory cache. Easyer to scale
Write though cache, first write cache then db
Write-back cache, first db then write cache

Master slave

Master write and propagte, read inconsitent only on slave. Like facbook where more read tha write

Distributed consensus

Multiple node agree for the final state. Odd number of nodes is better for votes

Consistency, use 3 databases in a triangular (Split-Brain). Because db can crash and network too.
Transaction: Only when transaction as been approved by all nodes, we may commit them -> consistent state committed, inconsistent transaction looking back and foward to propagate before writing
MVCC: protocol to get consistent read (heavy)
SAGA: very long transaction able to roll back in case of failure

Not Only SQL

Advantages are like json Vs comma separated values. Dynamic Vs Static. Array Vs List

SQL is said to be a relational Db.

Faster to read a full row (SELECT *) : all data is in one block
Scheme easily changeable: easy to add a col
Better scalable
Faster metric, analytics (average salary)
Bad for coherence ACID not garantee
- Slower read time (for a column)
- Relations are not implicits
- Joins are hard

Monolith Vs microservices

Advantages
- Less architecture
- Less moving parts
- Deployement, set up, test
- Faster, no RPC (if 2 micro services speak a lot, join them)
Drawbacks
- Integration of new team memeber, must be in context
- Deployelement must be more frequent
- Test complicates
- Single point of failuer
- More difficult to scale, more difficult monitoring, maybe the chat part limits, maybe the analytics
- More dificult to work in parrallel code in different files, machines, less dependencies
StackOverflow is a monolith

Publisher / Subscriber model

Pb: Request / Response has a stron coupling, client machine must wait a may timeout: failure latency.

Publisher / Subscriber model: there is a message broker between asker and responder that take the responsability to wait

Warning, no atomicity, poor consistency

Twitter is the perfect example: one publisher, many subscribers

Event driven (micro)services

The calling service is publishing events when he feels someone may need it. In a Request / Response services, he ask directly someone to do a job.

Avantages 1. Availability (! This always means inconsistency) 2. Easy Roll Back (Undo possible) 3. Replacement (due to easy rool back) 4. Transaction guarantee (can be send max or min 1 email) 5. Store intent

Drawbacks 1. Inconsistency 2. Gateway must store all responses 3. Lesser control from master -> dificult to trace data flow for devs

The publisher is sending an event to the event bus or message broker and who is interested can read from the event bus
Good in video games to get a consistent timeline of game: where was player2 when player1 shot, no when server received both info

Evernote

Mobile is the way in for new users.

Keywords

WebDav: file server
PubSub: message queue for Publisher Subscriber
GCS: Google Cloud Storage
RAID: Redundant Array of Inexpensive Disks
MySql: NoteStore
Lucene:
Graphite For logging with Slunk
Splunk ingestion pipeline.
Ansible: Automate deployment, configuration, instalation
Puppet: Declarative configuration language
Kubernetes: Orchestration framework

Experience first

Feature anti-pattern: 1. if a feature, you are thinking about must appear in preference (configurable): red alert. It is pointing the job to the user. 2. Disagree on the default value

Icons, and UX matters a lot when competing in the noise of conpetency (app, themes: food, music)

Design by Feature

Feature

Idea offline doc with zeal framework or webclipper

Web Clipper
Templates
Spaces
Integrations
Notes Sync
PDF & Doc Search
Search Handwriting
Document Scanning
Notebooks & Tags

Products

Evernote
Web clipper
Scannable
Skitch: draw on a image
Penultimate: Get presentation from hand written

Searching

Keyword
Tag