There are a couple of ways to do this.
The first is with Date Aggregation Operators, which allow you to dissect the "date" values in documents. Specifically for "grouping" as the primary intent:
db.collection.aggregate([
{ "$group": {
"_id": {
"year": { "$year": "$created_at" },
"dayOfYear": { "$dayOfYear": "$created_at" },
"hour": { "$hour": "$created_at" },
"interval": {
"$subtract": [
{ "$minute": "$created_at" },
{ "$mod": [{ "$minute": "$created_at"}, 15] }
]
}
}},
"count": { "$sum": 1 }
}}
])
The second way is by using a little trick of when a date object is subtracted (or other direct math operation) from another date object, then the result is a numeric value representing the epoch timestamp milliseconds between the two objects. So just using the epoch date you get the epoch milliseconds representation. Then use date math for the interval:
db.collection.aggregate([
{ "$group": {
"_id": {
"$subtract": [
{ "$subtract": [ "$created_at", new Date("1970-01-01") ] },
{ "$mod": [
{ "$subtract": [ "$created_at", new Date("1970-01-01") ] },
1000 * 60 * 15
]}
]
},
"count": { "$sum": 1 }
}}
])
So it depends on what kind of output format you want for the grouping interval. Both basically represent the same thing and have sufficient data to re-construct as a "date" object in your code.
You can put anything else you want in the "grouping operator" portion after the grouping _id. I'm just using the basic "count" example in lieu of any real statement from yourself as to what you really want to do.
MongoDB 4.x and Upwards
There were some additions to Date Aggregation Operators since the original writing, but from MongoDB 4.0 there will be actual "real casting of types" as opposed to the basic math tricks done here with BSON Date conversion.
For instance we can use $toLong and $toDate as new helpers here:
db.collection.aggregate([
{ "$group": {
"_id": {
"$toDate": {
"$subtract": [
{ "$toLong": "$created_at" },
{ "$mod": [ { "$toLong": "$created_at" }, 1000 * 60 * 15 ] }
]
}
},
"count": { "$sum": 1 }
}}
])
That's a bit shorter and does not require defining an external BSON Date for the "epoch" value as a constant in defining the pipeline so it's pretty consistent for all language implementations.
Those are just two of the "helper" methods for type conversion which all tie back to the $convert method, which is a "longer" form of the implementation allowing for custom handling on null or error in conversion.
It's even possible with such casting to get the Date information from the ObjectId of the primary key, as this would be a reliable source of "creation" date:
db.collection.aggregate([
{ "$group": {
"_id": {
"$toDate": {
"$subtract": [
{ "$toLong": { "$toDate": "$_id" } },
{ "$mod": [ { "$toLong": { "$toDate": "$_id" } }, 1000 * 60 * 15 ] }
]
}
},
"count": { "$sum": 1 }
}}
])
So "casting types" with this sort of conversion can be pretty powerful tool.
Warning - ObjectId values are limited to precision to the second only for the internal time value that makes up part of their data allowing the $toDate conversion. The actual inserted "time" is most probably dependent on the driver in use. Where precision is required, it's still recommended to use a discrete BSON Date field instead of relying on ObjectId values.
