時代とリサーチ
自分の会社が時代を通って進む理由、各時代のリサーチがその時代のトラフィックを必要とする理由、研究がインサイトを新しい道具に変える仕組み。
A Company grows through eight Eras, from the Garage to the Public Era, and each Era brings its problems and the Research to solve them together. New request kinds and failure modes arrive with the Era. You buy the tools with Insight, and your own traffic produces it, so you cannot research a problem you have not met.
What an Era changes#
You leave an Era by reaching its user milestone and taking (or declining) the funding round it offers. Cash, Burn, and funding explains that choice. Each new Era changes five things at once:
The request mix. New kinds of traffic arrive, such as writes and logins in Pre-seed, or search, uploads, and bots in Seed. Each kind loads your Design in its own way.
The market. Each Era has a larger market, which caps how many users you can win. Features can raise it.
The bar. The Apdex target rises each Era, so a Design that kept users happy in one Era may lose them in the next. From Series A, an availability SLO also counts every failed request.
Prices. A cost index and a salary index scale every price, so Burn stays a real share of revenue.
Incidents. From Pre-seed, incidents start, and later Eras allow more severe ones.
Here is the Seed Era's request mix:
- ページビュー26%
- API読み取り29%
- API書き込み12%
- ログイン6%
- 検索10%
- アップロード7%
- ボット10%
The Eras reference pages show each Era's milestone, market, targets, mix, and arrivals.
Insight comes in kinds#
Insight is the Research currency, and the Analytics Node makes it by sampling the Nodes you link into it. A sampled request gives Insight of its own kind: a sampled read gives read Insight, and a sampled search gives search Insight. Your Company keeps one pool for each request kind, plus a Failure pool that fills from requests that fail or frustrate users. Insight and observability explains the yield in detail.
The first time your Design serves more than half of a new request kind successfully, that kind's pool also gets a one-time gift of 20 Insight.
Recipes tie Research to traffic#
Each Research item has a recipe: the pools that pay for it. By default, an item's recipe is the request kinds of its Era. Some items ask for the kinds that show their problem: Caching asks for read Insight, and Connection pooling asks for write Insight. Half of every Reliability item's cost is Failure Insight, because you learn reliability by seeing failures.
So more page views never buy Series B Research. You must serve the new kinds first. Factory games use the same design with science packs, where each stage's production line pays for the next step.
No item opens before its Era, and none runs before the items it requires. Locked items of the next Era stay visible, so you can plan.
Making a kind with a Product lever#
A Study runs at the pace of its scarcest pool, so equal supply from every pool in its recipe is the target. The Era's mix works against you. Pre-seed gives writes several times the share of logins, and the default Pre-seed recipe pays for both, so every item on that recipe runs slowly on login. More samples raise every pool together. A Product lever changes the mix itself.
Each lever is a product change, such as a search box on every page or a nightly data export. It has three steps, and each step raises that kind's share of traffic. The other kinds give up the difference in proportion, so a lever moves demand from one kind to another instead of adding it. A step costs Dev Hours on your backlog and slows user growth. When you remove a step, half of its Dev Hours come back.
| ステップ | トラフィックの割合 | 開発時間 | 成長 |
|---|---|---|---|
| 1 | 12% | 24 | −1% |
| 2 | 18% | 80 | −3% |
| 3 | 24% | 240 | −7% |
The added requests earn what the traffic they displaced earned. They are marginal users of the same product, not new ones, so revenue per user stays where it was. What the lever buys is Research speed and a load you did not choose. Raising the login lever puts more password hashes on your App Servers, for example. That trade is why a lever on a kind you do not serve does nothing. A kind arrives with its Era and the Feature that brings it, and a lever only moves a share that already exists.
Find a lever in the Research panel. Open an item and read its Cost rows. The pool that limits the Study shows a card: its title says that kind's Insight is too low, it gives what the pool earns against what the Study needs, and its button names the step and its Dev Hours. Under the title, the card says what the step reaches in share of traffic and what it costs in user growth. When the Era is early, when a Venture caps the kind, or when a step already waits on the backlog, the button is greyed and the card gives the reason. The free Sandbox lists every lever in its current mix and sets a step at once, because it has no Team to build one.
Studies run one at a time#
A Research item runs as a Study, and one Study runs at a time. You can queue up to 8, reorder them, and switch which one runs. A Study keeps its progress when you switch away, but cancelling one loses its progress and the Insight it took. Its effect starts when it completes.
You can queue a locked item too, as Civilization lets you pick a distant technology and researches the path to it. Queue Node.js cluster mode without Node.js, and Node.js joins the queue ahead of it. The button says what it adds before you press it. Because the running Study must be one you can research, the queue holds an order: a Study never moves above one it needs, Study now brings an item's prerequisites to the front with it, and cancelling a prerequisite cancels what needs it. The cost is that a long chain fills the queue fast, and if the first Study in it stalls, everything behind it waits.
A Study draws Insight as it runs, not all at once. Each Era sets a base duration, from seconds in the Garage to an hour of sim time in the Public Era. Each second, the Study takes its share of the cost from every pool in its recipe. When a pool runs dry, the Study slows to the rate that pool earns, so the scarcest kind sets the pace. Traffic falls at night, so a Study that needs a quiet kind slows at night too.
You can start a Study with no Insight saved, and you cannot hoard much anyway. Each pool holds only what your Analytics retention window collects at the current rate, and never less than one Study's share. Studies run only while the Company is open; the clock stops when you close it.
Making Studies faster#
Two things raise a Study's speed:
A Spike. Put a Spike on a team's backlog, and that team works on the running Study instead of the product. While it does, the Study runs at twice the speed. Research competes with Features for engineers, as it does at real companies. See Team and velocity.
Study-speed Research. Trace search and Notebooks each add 25%.
Some Observability items also need breadth. Distributed tracing needs its kinds sampled at 3 Component types, because a trace needs more than one hop.
The tracks#
Research is grouped into ten tracks: Infrastructure, Data, Reliability, Observability, Performance, Runtimes, Frameworks, Security, Product, and Cost. Tracks are a way to browse, and an item in any track can require one in another. The Research reference lists every item with its Era, cost, and requirements.
Incidents also steer Research. When an incident ends, its postmortem names the first Research that would prevent it and that you can study now, and until you own it, that incident is likelier. How incidents work explains this loop.
Above the Company#
Eras and Research reset when a Company ends. Legacy, which you earn at an Exit, carries over. Why a Company ends explains that layer. To start, follow Earn Insight and research.