Can today’s record explain yesterday’s decision?

A customer asks why they received an overdue notice yesterday. Today their account says paid. Both facts can be true.

Give facts a time

Hickey distinguishes a fact from a later change. Overwriting the current status makes it hard to recover what the system knew when it sent the notice.

you can't update a fact is because you can't change the past

Versions can share structure

Keeping an old version need not mean copying the whole record. Persistent data structures share the parts that stayed the same; old and new readers each have a definite value.

the next version of the value shares a lot of structure with the prior version

What else is needed to reconstruct the state at the time?

Event sourcing reconstructs a past state by applying its ordered events to a projection. A record saying “just shut down” illustrates the remaining gap. In MIT’s reactor lecture, neutron-absorbing xenon builds up during operation, so the material present depends on earlier operation. A software projection can define how to replay discrete events; the physical process also evolves with time. A timestamped log is evidence of history. Reconstruction additionally needs the relevant initial conditions, inputs and transition rules.

Compare passages and timestamps

current state is a left fold of previous behaviors

Xenon-135 is one of the best neutron absorbers there is.

you'll reach the xenon equilibrium level where it will build to a certain level that will counteract the reactivity

Do two historical values need two copies of memory?

Hickey starts with locks that readers of shared data must remember. Immutable values reduce that coordination burden. Puente then opens the implementation: a small change in a persistent tree can copy the affected path while old and new versions share untouched branches. This explains why keeping a value may be cheap. The application still has to record which version actually triggered a particular reminder.

Compare passages and timestamps

have everybody remember what the locking model is

most of the data structure is shared between all these values

Is a recomputed July report the same report?

Hickey describes deriving prices from past sales. Kleppmann’s materialized view makes derived results cheap to read by maintaining them when underlying data changes. If both the orders and the pricing rule have since changed, today’s view answers a different question. Reproducing the quoted price requires the inputs and rule version used then, or the historical output itself.

Compare passages and timestamps

produces pricing information

query's already been done ahead of time

After the sale, which date values the proceeds?

This account of Japanese economic history separates asset sale, withdrawal date and price changes. A nominal amount can remain on the books while deferred access during inflation changes what it buys. Place pricing, access and the price index used for comparison on a timeline. A question about remaining wealth needs to specify which value on that timeline it asks for.

Compare passages and timestamps

Land reform during and after the war

they were not allowed to withdraw

The application still has to decide which versions to retain and record which version a decision used. Only then can supplying the old record and date reproduce the earlier answer.