TypedArray interface

[alippai]

Would adding a 32-bit date representation for Temporal.Date (eg. days since 1970-01-01 like Apache Arrow does) be in scope of this spec? Similarly 64-bit datetime for Temporal.DateTime could work.

Maybe this would help WASM interoperability too.

I know anybody can map the Date and DateTime to 32 and 64bit numbers resp., but adding the reference values would make sense. Adding a few extra methods to a Uint32Array/Float64Array subclass would make the developer experience better, too.

[justingrant]

Hi @alippai - There are two questions you're raising:

1. How should Temporal objects interoperate with WASM and other low-level interfaces?
2. How should Temporal objects be converted into a single number-since-epoch?

For (1), I'm not sure what thinking (if any) has been happening on that topic so I'll leave it to others. It's a good question.

For (2), a core concept in Temporal is the idea of "exact time" (measured in nanoseconds since epoch) and "local time" which displays human-readable values we think of as calendar dates and clock time. Converting exact time to a number is easy, using the Temporal.Instant type:

instant = Temporal.Instant.from('2020-01-01T00:00Z');
instant.getOffsetNanoseconds(); // this is a `bigint`
instant.getOffsetMilliseconds();  // this is a `number`
instant.toZonedDateTime('Europe/Budapest');  // coming soon
instant.toDateTime('Europe/Budapest');

Other than Temporal.Instant and Temporal.ZonedDateTime, the other Temporal date and time types types (Temporal.DateTime, Temporal.Date, Temporal.Time, Temporal.YearMonth, and Temporal.MonthDay) all are local time types. This means that they do not know their relationship to UNIX epoch. Additionally, Temporal.ZonedDateTime also stores a time zone and a calendar in addition to nanoseconds-since-epoch which means that it cannot be translated to a single number. Only Temporal.Instant can be losslessly serialized into a single 64-bit number.

Your question about Temporal.Date is an interesting one. The problem is that Temporal.Date stores (at least) four values internally: calendar (usually 'iso8601') and year, month, and day according to the ISO calendar. So there's no way to losslessly serialize a Temporal.Date into a single number without limiting it to the ISO calendar, which you can certainly do in userland.

[alippai]

@justingrant yes, it's indeed doable in userland. My idea was making a "fastpath" for the 80% of the use-cases both for:

1. the implementations (so they would just store it similarly as a 32-bit number and not converting it back and forth)
2. developers so when working across languages (either WASM or network) they wouldn't need to reimplement everything.

I wasn't aware of the different calendars, but having a default serialization at least of the "default" types (so Temporal.Date with iso calendar as int32 or Temporal.DateTime without timezone as something derived from int64 as it's not part of JSON).

Adding a recommendation to reduce the overhead of parsing/serializing could be a nice touch - for a single date it's not that important, but this is why I was asking about typed data. We don't have dataframe-like operations in JS, but there are use-cases where you want to keep a date array standard and compact, charting (x-axis) is one, often you pass thousands of values (coming from an API or generated in WASM).

My end-to-end example would be:

1. serialize a dataframe (timeseries data) using pyarrow
2. send it using HTTP or WebSockets
3. parse the table using Apache Arrow JS
4. convert it to a typedarray
5. pass the series to https://github.com/leeoniya/uPlot

Tldr; Highcharts, uPlot and others uses int timestamps for temporal data, this might be a use-case we want to support better

[ptomato]

We discussed this in the Oct. 8 meeting and have concluded that WASM code (or other code that requires a binary serialization of Temporal types) is not likely to have one binary serialization that we could prefer above all others. So, we will include an example of binary serialization in the cookbook, but we don't believe that there is one particular format that is used generally enough in interchange to be part of the API.

A possible route to achieving that might be to try to get an "ISO 8601 binary format" standardized.

[leeoniya]

(author of uPlot here)

there's definitely a void that needs to be filled in frontend applications that must deal with nanosecond resolution timestamps. right now all the options are rather awkward BigNum math or slow and reliant on string parsing like https://github.com/jcgertig/nano-date.

uPlot tries to minimize its use of Date objects in general and does a lot of the math externally on timestamps, only using Date for sanity-checking leap years, DST rollovers, etc.

[ptomato]

Cookbook example to be written after the freeze and before Stage 3