Why system design clicks when you watch the requests move
A system is a process, and moving pictures teach processes best one change at a time, explained aloud. What the best system design explainers do.
By the Chasca team 5 min read
System design is easier to understand when you can watch it run: a request leaving the browser, a cache answering before the database is asked, traffic moving away from a server that has just failed. The research on learning from pictures supports this, with conditions. The animation has to show the process itself, one change at a time, and the words have to arrive with the picture they describe.
Why does watching a request move help?
Because a system is a process, and processes are what animation shows well. Barbara Tversky, Julie Morrison and Mireille Bétrancourt called it the congruence principle: a picture works best when its form matches the idea it carries. A still architecture diagram shows what connects to what. Most of what system design is about happens over time: where a request goes, what waits in a queue, what breaks under load and what takes over.
On average, animations beat still pictures for learning. Tim Höffler and Detlev Leutner’s 2007 meta-analysis pooled 26 studies and found an advantage of d = 0.37, which held when the animation showed the content itself rather than decorating it (d = 0.40). Sandra Berney and Mireille Bétrancourt’s meta-analysis of 61 studies found a smaller overall benefit (g = 0.23) that grew when the animation came with spoken commentary (g = 0.34).
When does animation not help?
When it is fast, busy or decorative. The same 2002 paper warns that viewers take in continuous change as a series of steps, and that a picture only helps if it can be seen and understood accurately. Two findings show the way round it:
- One change at a time, with pauses. A meta-analysis of segmenting by Günter Daniel Rey and colleagues (56 studies, 7,713 learners) found that cutting a lesson into parts improved transfer (d = 0.36).
- Words beside what they describe, when they describe it. Paul Ginns’ meta-analysis of contiguity found large effects for putting related information together in space (d = 0.72) and in time (d = 0.78). On a diagram, that means a latency written on its arrow, not in a legend, appearing as it is said.
What do the best system design explainers do?
We measured six of the most-watched system design explainers on YouTube, from animated diagrams to whiteboard talks, a lightboard lesson and a one-hour interview walkthrough, using their subtitles and storyboard frames. The narrated, animated ones ran at about 141 to 166 words a minute. The ones with the most views a year changed something on screen every five to eight seconds and kept each frame sparse; the weakest stretches in the set held one picture still for minutes.
They also share a way of teaching:
- Start from a concrete problem with a number: a billion links, eleven thousand requests a second.
- Draw the simplest design that works, then show what breaks at scale.
- Fix it, and say what the fix costs: a cache can serve stale data, a second server needs spare capacity for failover, a rate limiter turns work away.
- Disturb the system and watch what moves: add a server, take one down, cut a link.
- Put the numbers on the diagram: a latency on an arrow, a load on a server.
- End by walking the final design, naming what each part does.
Why work the diagram out instead of drawing it?
Because a diagram that is slightly wrong teaches the wrong thing, and a system’s behaviour follows from its rules. Some places where it matters:
- Consistent hashing. In David Karger and colleagues’ 1997 paper, adding a server moves only the keys that now belong to it: about 1/(n+1) of them when one server joins n. A ring drawn by hand can easily show the wrong keys moving.
- Rate limiting. A token bucket, as in RFC 2697, has a rate and a burst size; a request that finds no token is refused, and on the web it gets 429 Too Many Requests. Which requests of a burst get through depends on exactly when they arrive.
- Back-of-the-envelope numbers. A billion requests a day is 1,000,000,000 ÷ 86,400, about 11,574 a second, and seven characters of Base62 give 62⁷, about 3.52 trillion codes. Numbers like these are better calculated than remembered.
- Latency. Jeff Dean’s well-known table of numbers puts a main memory reference at about 100 ns and a round trip within a data centre at about 500 µs. They are approximate and dated, but their orders of magnitude are why a cache in memory helps.
How do you make one with Chasca?
Chasca’s System design skill makes this kind of video from a system, a question or your own script:
- The diagram builds as it is said. Clients, load balancers, services, caches, databases and queues appear as the narration names them, laid out from left to right in the direction requests travel, so nothing has to be placed by hand.
- Requests move along the arrows, with latency and throughput on them, and a number counts to its new value when the design changes.
- The system’s behaviour is worked out: round robin skips a server that is down, a token bucket refuses what it has no tokens for, a queue counts what waits in it, and a hash ring gives each key to the next server clockwise.
- Sequence diagrams and hash rings in the same look, for protocols and for consistent hashing.
- The script is checked for factual mistakes before the video is made; if something needs a look, the video’s page says where.
- Any of 79 languages, with the technical terms engineers keep in English, and the diagram laid out right to left in Arabic.
The video at the top of this page is a one-minute example made with it: how a URL shortener resolves a link, adds a cache, scales out and keeps working when a server fails.
It explains systems in words and moving diagrams. It doesn’t read a picture of your diagram, and it never invents a product’s limits or prices.
Your first two videos are free, up to a minute each, watermarked and with no card required. Plans start at $9.99 a month (pricing).
Sources
- Tversky, B., Morrison, J. B., & Bétrancourt, M. (2002). Animation: can it facilitate? International Journal of Human-Computer Studies, 57(4), 247–262.
- Höffler, T. N., & Leutner, D. (2007). Instructional animation versus static pictures: A meta-analysis. Learning and Instruction, 17(6), 722–738.
- Berney, S., & Bétrancourt, M. (2016). Does animation enhance learning? A meta-analysis. Computers & Education, 101, 150–167.
- Rey, G. D., Beege, M., Nebel, S., Wirzberger, M., Schmitt, T. H., & Schneider, S. (2019). A meta-analysis of the segmenting effect. Educational Psychology Review, 31(2), 389–419.
- Ginns, P. (2006). Integrating information: A meta-analysis of the spatial contiguity and temporal contiguity effects. Learning and Instruction, 16(6), 511–525.
- Karger, D., Lehman, E., Leighton, T., Panigrahy, R., Levine, M., & Lewin, D. (1997). Consistent hashing and random trees. Proceedings of the 29th ACM Symposium on Theory of Computing, 654–663.
- Heinanen, J., & Guerin, R. (1999). RFC 2697: A Single Rate Three Color Marker. IETF.
- Nottingham, M., & Fielding, R. (2012). RFC 6585: Additional HTTP Status Codes. IETF.
- Dean, J. (2009). Designs, lessons and advice from building large distributed systems. Keynote, LADIS 2009.
Video transcript What the narrator says in the 1:11 explainer above.
A billion clicks a day: your short links must redirect without overwhelming storage.
Your browser calls an API. The API reads the destination from a database, then returns a redirect.
That averages over eleven thousand requests per second. Here, database lookups take eighty milliseconds. Storage becomes the bottleneck.
Add a cache for popular links. On a miss, read the database, then cache the destination.
Cache hits take two milliseconds here, skipping storage. Expire entries and invalidate changed links; cached destinations can become stale.
Add a load balancer, then another stateless API. Both share storage. Extra servers cost more.
What if one API fails? Health checks remove it; traffic moves to the survivor. Keep spare capacity.
Browsers request redirects. The balancer distributes clicks, APIs resolve links, the cache skips reads, and the database keeps mappings.
Questions people ask
On average, yes, when the animation shows the process itself. A meta-analysis of 26 studies found animations beat static pictures (d = 0.37), and a later one of 61 studies found the benefit grows when the animation comes with spoken commentary (g = 0.34). Animations that are too fast or too busy help less.
It starts from a concrete problem with a number, builds one diagram in the order the voice names its parts, shows requests moving and failures happening, says what every fix costs, and ends by walking the final design. The most-watched animated explainers we measured change something on screen every five to eight seconds.
Only the keys that now belong to the new server, about 1/(n+1) of them when one server joins n. That follows from the monotonicity and balance properties in Karger and colleagues' 1997 paper. With hash mod n, almost every key moves.
Chasca's System design skill makes one from a system, a question or your own script: the diagram is laid out for you as the narration names each part, requests move along the arrows, and round robin, token buckets, queues and hash rings are worked out rather than drawn.
Yes. The skill narrates in any of Chasca's 79 languages, keeps the technical terms engineers use in English, and lays the diagram out right to left in Arabic.