[ ]productinterview
Run a free mock

system design · hard

"Design Instagram's system" PM interview answer

Design Instagram's system.

Updated Jun 2026 Calibrated to the strong-hire bar

This question tests whether you understand the product consequences of architectural choices, not whether you can draw a distributed systems diagram. A weak answer narrates boxes and arrows. A strong one names who gets hurt when you pick the wrong fan-out model, and why Instagram’s actual 2026 architecture can no longer be described as a social graph problem at all.

Scope-set before you design

Open with a 90-second constraint: “I’ll focus on the core Instagram experience in 2026: media creation, the ranked feed, and the follow graph. I’ll skip Stories, DMs, and Live unless you want to go there.” Then name the three hardest PM-level problems you intend to address: (1) feed quality versus latency, (2) creator fairness at asymmetric graph scale, and (3) the structural dependency between the ranking system and the ad auction. If you only address (1), you’re answering the 2013 version of this question.

The social graph is not the hard part

Instagram’s follow graph is directed and asymmetric: following someone does not mean they follow you back. The core traversal is simple: “give me the 500 accounts this user follows.” That maps naturally to an adjacency list, not a full graph database. Multi-hop queries (friends-of-friends) are rare enough that you can compute them offline and cache the results.

The hard part is the edge case that becomes the norm at Instagram’s scale: a Cristiano Ronaldo post to 600 million followers. Fan-out-on-write (the push model) precomputes each follower’s feed via a message queue at upload time. It gives fast reads, but a single celebrity post triggers 600 million write events. That is not a backend scaling footnote; it is a creator experience problem. High-follower creators on a pure push model see slower post-to-feed-appearance latency because the system is throttled to manage fan-out load. Instagram solves this with a hybrid: fan-out-on-write for accounts under roughly 1 million followers, fan-out-on-read (pull model: assemble feeds at request time) for high-follower accounts. The PM-level insight is that Instagram patches the creator experience gap by showing the post on the creator’s own profile immediately while async fan-out completes.

Scale context: 500 million DAU, 100 million posts per day across photos, videos, and Reels, roughly 100TB of new media ingested daily.

Media storage: the tradeoff most candidates miss

Clients upload directly to pre-signed object storage URLs (S3-equivalent), bypassing the application server entirely. This is good for upload latency. The tradeoff: the application server is not in the media path, which means you cannot do synchronous content moderation before the file lands in storage. You need async moderation with a quarantine state: media is ingested, a moderation pipeline runs, and content stays in a non-public state until it clears. Some violating content will be briefly stored before detection. Name this explicitly. Interviewers notice when candidates present CDN-via-S3 as a clean win without acknowledging the moderation timing gap.

CDN cache hit rate is the primary lever for read latency on media. Sub-500ms feed load is the standard target; the CDN is what makes that achievable at global scale, not the application layer.

The 2026 feed is not the social graph

This is the single most important update to give any interviewer who frames this as a feed-from-follows question: Instagram’s feed in 2026 is approximately 30% content from accounts the user does not follow, ranked by an ML recommendation model. The social graph is one ranking signal, not the architecture.

The system now maintains two parallel graphs: the social graph (who you follow) and an interest graph (engagement history, content embeddings, affinity by content type). The recommendation model runs at feed-assembly time and consumes the majority of the latency budget. Ranking factors, in rough descending weight: engagement depth signals (save and share outweigh like, which outweighs view), recency, relationship strength (DM history and mutual follows get a boost over asymmetric follows), and content-type affinity (a user who watches Reels to completion but skips static posts gets a different mix than one who does the reverse).

The PM-level tension here: every point of feed surrendered to the interest graph improves discovery and aggregate engagement, but users lose the intuition for why they’re seeing something. Above a threshold, that erodes trust and increases unfollows. You should name this explicitly, and propose a UX mechanism (topic tags, “suggested for you” labels) that surfaces the recommendation rationale without alarming users about what the system infers about them.

The ad auction is not separable from feed generation

Instagram’s revenue model means 2-3 feed slots per session are reserved for ads. The ad auction must resolve within the feed latency budget, which is roughly 50-100ms for ad retrieval. The practical PM design choice: use a lightweight, partially cached auction result rather than a fully fresh real-time auction on every feed load. This is a deliberate product-business tradeoff: slightly suboptimal ad targeting in exchange for acceptable feed load time. A candidate who treats the ad slot as a cosmetic addition to the feed has missed that the ranking system and the monetization system are the same system with shared latency constraints.

Structure a strong answer

strong

"Let me scope this to the core Instagram experience in 2026: media creation, the ranked feed, and the follow graph. I'll skip Stories and DMs unless you want to go there. The three PM-level problems I'll focus on are feed quality versus latency, creator fairness at asymmetric graph scale, and the dependency between the recommendation model and the ad auction.

On the social graph: directed, asymmetric, stored as an adjacency list. The core query is always 'give me the accounts this user follows.' That doesn't require a full graph DB. The hard case is celebrities with 600M followers. Pure fan-out-on-write would trigger 600M queue events on a single Ronaldo post. So Instagram uses a hybrid: push fan-out for normal accounts, pull fan-out on read for accounts above roughly 1M followers. The PM tradeoff is that high-follower creators see slower feed distribution latency. Instagram patches this by surfacing the post on the creator's own profile immediately while async fan-out runs in the background.

On media storage: clients upload directly to pre-signed object storage URLs, bypassing the app server. Good for upload latency, but it means moderation is asynchronous. You need a quarantine state: content ingested, moderation pipeline runs, content remains non-public until it clears. Some violating content will be stored before detection. Name this, because it's a product tradeoff, not just an ops detail.

On the feed in 2026: I would push back on framing this as a social-graph fan-out problem. Roughly 30% of Instagram's feed is content from accounts the user does not follow. The system maintains a social graph and a separate interest graph built from engagement history and content embeddings. The recommendation model runs at feed-assembly time and consumes most of the latency budget. The ranking factors that matter most are engagement depth signals (saves and shares), recency, relationship strength, and content-type affinity. The PM judgment call: how much of the feed to give to the interest graph versus the social graph. More interest-graph content means better discovery and higher session engagement, but above a threshold users lose the sense of why they're seeing something, which erodes trust. I'd propose topic labels and 'suggested for you' signals to maintain transparency without alarming users about what the system infers.

On ads: 2-3 feed slots per session are reserved for ads. The auction must clear within roughly 50-100ms of the feed latency budget. I'd use a cached auction result, not a fully real-time auction on every load. This accepts slightly suboptimal ad targeting in exchange for predictable feed load time. The ranking system and the ad auction share the same latency envelope, so you can't design them independently."

weak

"I'd add a load balancer, application servers, Postgres for metadata, S3 for media, Redis for caching, and a CDN. For the feed, I'd use fan-out-on-write so each user's feed is precomputed. I'd add monitoring and alerting." This is engineer-cosplay. It narrates the request path without connecting any choice to a user or business outcome. It picks chronological fan-out-on-write without acknowledging that Instagram's feed is not chronological and that pure fan-out-on-write collapses at celebrity scale. It treats the ad slot as decoration. It gives no PM judgment: no tradeoffs named, no user segments considered, no failure modes discussed. An interviewer at Meta will immediately ask "what happens when Ronaldo posts?" and this candidate will have nothing to say.

What the interviewer is actually checking

At the PM level, system design questions are tests of product reasoning, not infrastructure recall. The question is not “do you know what a CDN is?” It is: “can you trace the product consequence of each architectural choice?” Fan-out-on-write is not interesting because it’s a distributed systems pattern; it’s interesting because it creates a two-tier creator experience between normal accounts and high-follower accounts. The media pipeline’s async moderation is not interesting as a queue architecture; it’s interesting because some fraction of violating content will be briefly accessible before detection, which is a product safety risk you must acknowledge and design for.

The 2026 reframe to land before you close: Instagram’s system is no longer primarily a social-graph distribution problem. It’s a recommendation system problem that uses the social graph as one signal among several. The viable and lovable tension sits in the interest graph: more recommendation means better discovery and higher ad revenue viability, but only if users feel anticipated rather than surveilled. The system design embeds that product question in every architectural choice about how much weight the ranking model gives to interest-graph signals versus follow-graph signals.

Asked at

Related