[
https://issues.apache.org/jira/browse/CALCITE-1237?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15290715#comment-15290715
]
Fabian Hueske commented on CALCITE-1237:
----------------------------------------
I think the behavior of {{session}} (with or without all grouping columns)
would be rather like a ranking function (such as {{NTILE}}) than like an
aggregate function. {{session}} would return a {{window_id}} for each record in
a partition of records and not only one {{window_id}} for each detected session
in the partition.
I think if we want to be strict, the ranking function would need to be followed
by an {{OVER}} clause that defines partitioning (on the additional grouping
columns) and the order (on a monotone attribute). Maybe this can be relaxed we
implicitly define the sorting on the first parameter and the partitioning on
the remaining column parameters of {{session}}.
Also, not all ranking functions can be allowed. {{NTILE}} requires to know the
number of elements in a partition which is infinite for streams.
> Session windows for streaming SQL
> ---------------------------------
>
> Key: CALCITE-1237
> URL: https://issues.apache.org/jira/browse/CALCITE-1237
> Project: Calcite
> Issue Type: Bug
> Components: stream
> Reporter: Julian Hyde
> Assignee: Julian Hyde
>
> A session window is a collection of rows whose key values, when sorted, have
> a gap of at most N.
> Q1. Should "at most" be "less than"?
> The key type can be any type that has a minus operator, that is, numeric and
> date-time.
> I propose the following syntax: {{session(key [, ...]*, interval)}}. For
> example:
> {code}
> select stream session(rowtime, productId, interval '5' second),
> productId, count(*) as c
> from Orders
> group by session(rowtime, productId, interval '5' second),
> productId
> {code}
> to find bursts of orders for the same product where consecutive orders are no
> more than 5 seconds apart.
> The first key column {{rowtime}} defines the session and must be of
> numeric/date-time type, and must have monotonicity or similar in order for
> the query to make progress; the other key columns (in this case
> {{productId}}) can be of any type; the last column is the interval, and must
> be constant.
> The {{session}} function returns the key value at the start of the window.
> Unlike the {{hop}} function, each row belongs to precisely one window. But
> {{session}} is not a true function, because its value depends on the records
> flowing in the stream.
> Q2. If {{session}} is used, should we allow order-dependent aggregate
> functions such as {{first_value}}?
> Q3. Should we allow {{session}} as a windowed aggregate function?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
