OpenAI ne bada karan Git backend ke liye Pando vikas kiya hai

OpenAI ne Pando vikas kiya hai, jo Git repositories ke liye ek naya backend hai jo OpenAI ke monorepo ke badhte hue pariman aur mahatva ko sambhalne ke liye banaya gaya hai. OpenAI ke anusar, is system ko Codex dwara sambhavit uchch-samavayi kaam ko support karne ke liye banaya gaya hai, bina maujooda engineering workflows ko rukaye ya team ke Git istemal ke tareeke mein koi badlav kiye.

Scalability aur Performance

OpenAI ke anusar, Pando ko do mukhya lakshyon ko palne ke liye banaya gaya hai: bahut se shakhaon mein samayein (concurrent) pushes ke liye uchch-samavayi likhne ke liye horizontal roop se scalable likhne ki prapti, aur padhne ki tezta ko uchch karke clients ko aise data download karne ki ijazat dena jisse available network bandwidth ko pura kiya ja sake.

Production mein, OpenAI ke anusar Pando ne neeche diye gaye lakshyon ko prapt kiya hai:

• Median push latency ko lagbhag 5 guna kam kiya.
• Rozana lagbhag ek petabyte Git data pradan kiya.

Technical Architecture

In lakshyon ko palne ke liye, Pando ek aaramda samajh (relaxed consistency model) jise causal consistency kehte hain, ka upyog karta hai. Is tareeke se, alag-alag shakhaon (unrelated branches) mein hone wali updations ko samayein (parallel) aage badhaya ja sakta hai kyunki har shakha ki updations ko alag-alag coordinate kiya jata hai. OpenAI ke anusar, iska matlab hai ki koi padhne wala (reader) jo kai shakhaon ko dekhta hai, wo unhe alag-alag samay ke bhaag mein dekh sakta hai.

Purane Git service patterns ke vibhinn, Pando neeche diye gaye vidhiyon ka upyog karta hai:

• Likhne ki Optimization: Pando aane wale packfiles ko unpack karke local repository mein dalne ke bina, unhe blob storage mein stream karte waqt validate karta hai. Yeh ek multi-key transaction ka upyog karta hai metadata store mein ref versions ko record karne ke liye, jise OpenAI ke anusar $O(1)$ metadata reads aur writes ki zarurat hoti hai.
• Padhne ki Optimization: Pando har client ke liye alag-alag packfiles banane ke bajaye fetches ke dauran existing packfiles ko wapas istemal karta hai. Isse server ko bahut kam computation ke saath kai clients ko same bytes pradan karne ki ijazat milti hai.

Existing Methods se Samanya

OpenAI ne Git ko scale karne ke do prachin tareekon ki nirnay kiya: three-phase commit (3PC) aur write-ahead log (WAL) based designs.

OpenAI ke simulations aur testing ke anusar:

• Three-phase commit (3PC): Is vidhi mein coordinator aur replicas ke beech kai rounds communication ki zarurat hoti hai, jise OpenAI ke anusar load ke dauran disks saturate hone par likhne ki latency mein bahut zyada badhavati hoti hai.
• Write-ahead log (WAL): Is vidhi blob storage mein ek shared manifest ka upyog karta hai updates ko order karne ke liye. OpenAI ke anusar, in systems mein padhne ki latency bahut zyada hoti hai kyunki servers journal se padhne mein peeche reh jaate hain.

OpenAI ke anusar, Pando load ke dauran performance par asar na padhne ke saath kai hazaron clients ko samayein (concurrently) push karne ki suvidha deta hai.

स्रोत (अंग्रेज़ी में)

• openai.com: Scaling Git at OpenAI | OpenAI

यह मूल अंग्रेज़ी लेख का एआई द्वारा किया गया अनुवाद है, जिसे स्वचालित रूप से जाँचा गया है ताकि यह मूल लेख जैसी ही बात कहे। स्रोतों की जाँच अंग्रेज़ी में की गई थी।