Cloud Firestore: Storing and querying for today's date over multiple UTC offsets?

Question

I'm writing a web app using Firestore that needs to be able to show "Today's Popular Posts" and I'm having trouble getting the queries right when considering users in different timezones.

The dates are stored in the DB as UTC 0, then adjusted to the current user's UTC offset in the client via Moment.js. This works correctly.

When adding a new post I use firebase.firestore.FieldValue.serverTimestamp() to store the current server timestamp in a field called timestamp, like so:

const collectionRef = db.collection('posts');
collectionRef.add({
  name: "Test Post",
  content: "Blah blah blah",
  timestamp: firebase.firestore.FieldValue.serverTimestamp(),
  likeCount: 0
});

Then on the server I have a Cloud Function that runs on create and adds another field to the document called datestamp which is the the UTC 0 timestamp, but adjusted so that the time is the beginning of the day. The function looks like this:

exports.updatePostDate = functions.firestore
  .document('posts/{postID}')
  .onCreate((event) => {
    const db = admin.firestore();
    const postRef = db.doc('post/'+event.params.postID);
    const postData = event.data.data();

    const startOfDay = moment(postData.timestamp).startOf('day').toDate();

    return postRef.update({
      datestamp: startOfDay
    });
  });

Storing a timestamp where the time is always the beginning of the day enables me to write a query like this for finding all posts and ordering by popularity on a given day:

const startOfDayUTC = moment.utc().startOf('day').toDate();
const postQuery = db.collection('posts')
                    .orderBy('likeCount', 'desc')
                    .orderBy('timestamp', 'desc')
                    .where('datestamp', '==', startOfDayUTC)
                    .limit(25);

The problem is, depending on the user's UTC offset, this can display posts with two different dates when parsing the post's timestamp field. So even though the query is correctly fetching all the posts where the datestamp is say, 2018-01-30T00:00:00Z, the timestamp's date might not be the same once parsed. Here's an example of two posts:

Post 2:
likeCount: 1
timestamp (UTC 0): 2018-01-30T06:41:58Z
timestamp (parsed to UTC-8): 2018-01-29T22:41:58-08:00
datestamp (UTC 0): 2018-01-30T00:00:00Z

Post 1:
likeCount: 0
timestamp (UTC 0): 2018-01-30T10:44:35Z
timestamp (parsed to UTC-8): 2018-01-30T02:44:35-08:00
datestamp (UTC 0): 2018-01-30T00:00:00Z

So you can see, while the posts have the same datestamp, after adjusting the timestamp to the local UTC, the timestamp fields can end up being on two different days.

If anyone has a solution to this I would be very grateful.

Answer 1

I think it is better to avoid functions in this case as you can perform compound queries now. You can simply use

query.where(date > lastMidnight).where(data < now).get().then(...)

so to speak to limit data which only belongs to one day and try to keep all your time variables in UTC 0 and just find the start point and the current time both client side and convert them to UTC0.

//get local time from midnight to now (local)
const now = new Date();
const lastMidnight = now.setHours(0,0,0,0);

//then convert those to UTC0 to pass on in your query to firestore
const lastMidNightUTC = new Date(lastMidnight + now.getTimezoneOffset() * 60000).toString();
const nowInUTC = new Date(now + now.getTimezoneOffset() * 60000).toString();

and you can get your data (remember you need to make an index or just run the query once and firebase SDK will generate a link to create the index in dev tools -> console , for you)

    query.where(date > lastMidNightUTC).where(data < now).get().then(...)

Answer 2

I came up with a solution that I'm really not happy with... But it works!

The problem is fundamentally one post can be on more than one date, depending on the user's location. And since for this case we also want to order by a field other than timestamp we can't use a range query to select posts on a given date, because your first .orderBy must be on the field you're using a range query on (see Firestore docs).

My solution is to map localized datestamps to their corresponding UTC offset. The object contains every UTC offset as a key, and the post's datestamp in that offset's time.

An example post looks like this:

posts/{somepostid}
{
  name: "Test Post",
  content: "Blah blah blah",
  timestamp: Mon Jan 29 2018 21:37:21 GMT-0800 (PST),
  likeCount: 0,
  utcDatemap: {
    0: "2018-01-30,
    ...
    -480: "2018-01-29",
    ...
  }
}

The field utcDatemap is the the one we use in our queries now:

const now = moment();
const datestamp = now.format("YYYY-MM-DD");
const utcOffset = now.utcOffset();
const utcDatemapField = 'utcDatemap.'+utcOffset;

const postQuery = db.collection('posts')
                    .orderBy('likeCount', 'desc')
                    .orderBy('timestamp', 'desc')
                    .where(utcDatemapField, '==', datestamp)
                    .limit(25);

Now posts can show up on two different days, depending on where the user is querying from. And we can still convert the regular old timestamp to the user's local time on the client.

This is definitely not an ideal solution. For the above query I needed to create composite indexes for every single key in utcDatemap. I'm not sure what the rules of thumb are with composite indexes, but I'm thinking having 39 indexes for something simple like this is probably not great.

Additionally I checked it out using the roughSizeOfObject function from thomas-peter's answer on this post and the utcDatemap object, with all it's string datestamps clocked in at roughly 780 bytes, and it's not like 0.78kb is a lot, but you do need to be mindful of how much data you're transferring with a service like Firestore (0.78kb is a lot for a date).

Answer 3

I'm learning/reading up on Firestore and have Google'd to see how it deals with times, so discount my answer appropriately.

It looks as though Firestore converts times to UTC and stores them as its own Timestamp datatype. If so, then it's critical to know that this is a destructive conversion.

Though UTC is useful for comparing instants in time, it means that the wall-clock time as observed by the app user is lost forever. Some countries like the UK are in one of two timezones during the year, Daylight Savings Time and British Summer Time.

You can convert back to the user's observed time, but the problem is that the rules change over the years. You'd have to keep a record of all the different rule changes for all the timezones and countries of the world.

What was the time at the time?

The question is, what time did the user think an event happened ...at the time. This can have legal ramifications. You may need to go back through historic data to prove a person acted at a certain time as they observed it.

The solution is to capture the user's observed offset in an additional field. That way, you can always use this to convert.

Regarding the OPs problem, this seems somewhat philosophical for a web app. Does "today" mean the event, such as a post, must have happened within the user's Monday? Or just posts on today's date? Or posts within the last 24h?

An important thing to remember is that dates are the same all around the world, even when they begin and end at different instants.

What's Elvis got to do with all this?

Christmas Day is 25th everywhere. If I say something happened on Christmas Day and I'm in the USA, and then someone in Australia wants to see all the world's posts made on Christmas Day, then they need to query for posts where observedDate == 25th Dec.

Think about it. Such posts were all made on Christmas Day, even though it might have been Boxing Day for me in England at the instant that they posted.

Elvis died on 16th August. In the UK our radio stations don't all wait until it's the 16th in the timezone of the place of his death to start playing his records.

Another interesting one is whether something happened in Q1 or Q2 of a company's reporting year. Is a sale recognised as on the date at the point-of-sale in the store in New York, or in the database in LA?

The observed date is interesting.

My point is, think deeply about this and store both a normalised instant like UTC, but also the user's observed date, time and offset. Then you have all the information you'll need for the future.

Finally, consider adding observed year or week numbers, or day ordinals, H1/H2 for financial data, since it can be super useful for composing rapid queries, depending on your use-cases.