Chapter 2 · Hinglish · Sunne-wala Sabak
Envelope Estimation — Napkin Maths
Capacity, latency, storage aur bandwidth ka jaldi wala andaaza. Niche "Play all" dabaiye aur browser pura chapter padh kar sunayega. Kisi bhi section par "Suniye" se wahin se shuru karein.
▶ Companion slides kholेंTip: jo voice sabse natural lage wahi chuniye — ek Hindi (hi-IN) ya Indian-English (en-IN) voice aam taur par best chalti hai.
Estimate kyun karein Slide 2
Chalo chapter do shuru karte hain, back-of-the-envelope estimation. Yeh ek forecast nahi hai, yeh ek feasibility check hai. Yeh seconds mein bata deta hai ki design roughly sahi hai ya nahi, aapko ek box chahiye ya ek hazaar, aur asli bottleneck kahan hoga. Poora skill ek hi move hai jo baar baar karte hain. Ek vague problem ko chhote chhote multipliers ki chain mein todo, aur phir number ko khulkar round kar do. Interview mein number se zyada yeh maayne rakhta hai ki aap us number tak bina ghabraye pahunch paate ho ya nahi.
Powers of two Slide 3
Pehla building block, powers of two. Inhe ek baar yaad kar lo to storage ka maths dimaag mein hi ho jaata hai. Mental math ke liye, do ki power das ko lagbhag ek hazaar maano, do ki power bees ko lagbhag das lakh, aur do ki power tees ko lagbhag ek arab. Teen aise numbers multiply karne par bhi error sirf kuch percent aata hai. Ek choti baat, binary aur decimal units thode alag hote hain. Ek kilobyte ek hazaar bytes hota hai, par ek kibibyte ek hazaar chaubees bytes. Napkin math ke liye dono barabar, par disk aur billing ke liye nahi.
Latency ladder Slide 4
Doosra building block, latency ladder. Yeh orders of magnitude yaad rakhne layak hain. Sabse important reflex yeh hai. RAM, SSD se lagbhag sau guna fast hai. SSD, spinning disk se lagbhag sau guna fast hai. Ek hi data center ke andar round trip RAM aur SSD ke beech hota hai. Aur cross-region har hop par tens of milliseconds jodta hai jise aap optimise nahi kar sakte. Light fibre mein roshni ki do-tihai speed se chalti hai, aur yahi har cross-continent hop ka floor hai.
Availability ke nines Slide 5
Teesra, availability ke nines. Har extra nine lagbhag das guna mushkil hota hai. Do nines, yaani 99 percent, ka matlab hai saal mein lagbhag paune chaar din downtime. Teen nines, yaani 99.9 percent, ka matlab saal mein lagbhag nau ghante. Ek dhyaan ki baat, nines neeche ki taraf multiply hote hain. Agar ek request teen services se guzarti hai aur har ek 99.9 percent par hai, to effective availability lagbhag 99.7 percent reh jaati hai. Har synchronous dependency aapke headline number par tax lagati hai.
QPS — users se load tak Slide 6
Ab traffic. QPS, yaani queries per second. Pehle user count lo, phir behaviour, phir peakiness. DAU ko per-user daily actions se multiply karo, aur ek din ke seconds se bhaag do. Yaad rakho, ek din mein lagbhag chaurasi hazaar seconds hote hain, yaani lagbhag ek lakh. Isse average QPS milta hai, jo aapke steady-state fleet ka size batata hai. Phir average ko do se paanch guna karo, yeh peak QPS hai, jo worst case headroom batata hai. Aur reads aur writes ko hamesha alag alag estimate karo.
Storage ke chaar multipliers Slide 7
Storage ke chaar multipliers hain, isi order mein. Items per day, multiply average size, multiply teen sau pachsath din, multiply replicas, multiply retention years. Inme se ek bhi chhoda to answer ek order of magnitude galat ho jaata hai. Replication aam taur par teen guna hota hai, indexes 30 se 80 percent jodte hain, aur logs aur backups upar se aur. Aur growth multiply hoti hai, linearly add nahi hoti.
Bandwidth aur bits-vs-bytes Slide 8
Bandwidth simple hai. Egress matlab read QPS guna average response size. Ingress matlab write QPS guna average request size. Par ek catch hai, bits aur bytes mat confuse karna. Das gigabit per second ka matlab sirf sava ek gigabyte per second hai, aur overhead ke baad usse bhi kam. Yeh aath guna ki galti hai, sabse common. Aur cloud mein egress hi aapke bill par dikhta hai. Ek availability zone ke andar bandwidth lagbhag free hai, zones ke beech mehnga, aur regions ke beech ya internet par sabse mehnga.
Example — social feed Slide 9
Ek example lete hain, ek microblog jise hum Quill kehte hain. Maan lo chaar crore daily active users hain. Reads bahut zyada hain, writes kam. Saara maths karne par storage chhota nikalta hai, lagbhag chaudah terabyte per year. Par peak reads bahut zyada, lagbhag bayalees hazaar per second. To yahan asli design driver storage nahi, balki read fan-out hai. Celebrity authors ke liye ek hot cache chahiye.
Example — photo app Slide 10
Doosra example, ek photo app jise hum Pebble kehte hain. Yahan shape ulta hai. Photos storage aur egress dono ko dominate karti hain. Storage petabyte-scale mein chala jaata hai, lagbhag saadhe aath petabyte per year. Aur peak egress itna zyada ki ek single origin se nahi chal sakta, lagbhag pachpan gigabit per second. Iske liye ek CDN with high hit ratio zaroori hai. Upload path chhota hai, kuch dozen workers kaafi.
Common galtiyan Slide 11
Ab common galtiyan. Ek, bits aur bytes confuse karna, jo aath guna ki galti hai. Do, average ke liye size karna, peak ke liye nahi. Teen, replication aur indexes bhool jaana. Chaar, growth bhool jaana. Paanch, false precision, jaise saat digit ka QPS likhna jab ek rounded number zyada honest hai. Aur chhah, long tail ko ignore karna, kyunki kuch celebrity ya viral items lakhon normal users se zyada load la sakte hain.
Rules of thumb Slide 12
Kuch rules jo aapko honest rakhte hain. Liberally round karo, ek significant figure kaafi hai. Single number ki jagah ek range do, jaise aath se baarah petabyte. End to end sanity check karo, agar answer pagal lage to ruk jao. Aur apni assumptions likh do, kyunki estimate ki asli value final number nahi, balki assumptions ka trail hai.
Paanch principles Slide 13
To paanch principles yaad rakho. Ek, decompose karo phir multiply karo. Do, average fleet ka size batata hai aur peak headroom ka. Teen, powers of two aur latency ladder ko reflex bana lo. Chaar, replication, indexes, aur growth optional nahi hain, yeh aam taur par raw data ka paanch se das guna hote hain. Aur paanch, round karo, range do, aur assumptions likho.
Aage kya
Bas, yahi hai chapter do. Ek kaam karo, koi bhi product chuno jise aap use karte ho, aur use loud bolkar estimate karo. DAU se actions, phir QPS, phir storage, phir egress, har step par hard round karte hue. Method ko reflex bana lo, yahi sabse zaroori hai. Koi multiplier ya unit confuse kare to mujhse pooch lena. Agle chapter mein hum ek rate limiter design karenge.