Elastic aggregation to identify period A vs B percentage increases - elasticsearch

Question

I have some daily sales data indexed into Elasticsearch. I successfully run a number of aggregations to identify top sellers across a date range etc.
I am now trying to write a single query to do the following:
Identify Top n sellers over a date range (Period A)
Take the results of Period A and sum sales for these products over second date range (Period B)
Compare sales in period A to Period B and identify those with percentage increases above X%.
My attempt so far:
{
"query": {
"bool": {
"filter": [
{
"range": {
"date": {
"gte": "2017-10-01",
"lte": "2017-10-14"
}
}
}
]
}
},
"size": 0,
"aggs": {
"data_split": {
"terms": {
"size": 10,
"field": "product_id"
},
"aggs": {
"date_periods": {
"date_range": {
"field": "date",
"format": "YYYY-MM-dd",
"ranges": [
{
"from": "2017-10-01",
"to": "2017-10-07"
},
{
"from": "2017-10-08",
"to": "2017-10-14"
}
]
},
"aggs": {
"product_id_split": {
"terms": {
"field": "product_id"
},
"aggs": {
"unit_sum": {
"sum": {
"field": "units"
}
}
}
}
}
}
}
}
}
}
Although this outputs results for two periods, I don't think this is quite what I want as the initial filter is running from Period A start date to Period B end date and I think summing results for that range instead of Period A only. I also don't get the % comparison, I would probably do this at my application level, but I understand could be handled with a scripted Elastic query?
It would be especially awesome if instead of top n results in period A, I could set a sales threshold of say 1,000 sales.
Any pointers would be much appreciated. Thanks in advance!
Currently running Elastic 5.6

Answer 1

{
"query": {
"bool": {
"filter": [
{
"range": {
"date": {
"gte": "2017-10-01",
"lte": "2017-10-14"
}
}
}
]
}
},
"size": 0,
"aggs": {
"data_split": {
"terms": {
"size": 10,
"field": "product_id"
},
"aggs": {
"date_period1": {
"filter": {
"range": {
"date": {
"gte": "2017-10-01",
"lte": "2017-10-07"
}
}
},
"aggs": {
"unit_sum": {
"sum": {
"field": "units"
}
}
}
},
"date_period2": {
"filter": {
"range": {
"date": {
"gte": "2017-10-08",
"lte": "2017-10-14"
}
}
},
"aggs": {
"unit_sum": {
"sum": {
"field": "units"
}
}
}
},
"percentage_increase": {
"bucket_script": {
"buckets_path": {
"firstPeriod": "date_period1>unit_sum",
"secondPeriod": "date_period2>unit_sum"
},
"script": "(params.secondPeriod-params.firstPeriod)*100/params.firstPeriod"
}
},
"retain_buckets": {
"bucket_selector": {
"buckets_path": {
"percentage": "percentage_increase"
},
"script": "params.percentage > 5"
}
}
}
}
}
}
And a full test data in this gist.
The result of this aggregation is giving you this:
"aggregations": {
"data_split": {
"doc_count_error_upper_bound": 0,
"sum_other_doc_count": 0,
"buckets": [
{
"key": "A",
"doc_count": 6,
"date_period1": {
"doc_count": 3,
"unit_sum": {
"value": 150
}
},
"date_period2": {
"doc_count": 3,
"unit_sum": {
"value": 160
}
},
"percentage_increase": {
"value": 6.666666666666667
}
},
{
"key": "C",
"doc_count": 2,
"date_period1": {
"doc_count": 1,
"unit_sum": {
"value": 50
}
},
"date_period2": {
"doc_count": 1,
"unit_sum": {
"value": 70
}
},
"percentage_increase": {
"value": 40
}
}
]
}
}
The idea is that you use two filter type of aggregations for the two date intervals. And for each you calculate a sum. Then, using a third aggregation of type bucket_script you calculate the percentage increase (note, though, that it will be a negative number of there is a decrease in sales for example).
Then, using yet another aggregation - of type bucket_selector - you keep the product_ids where the percentage is larger than 5%.