I think my previous comment about maintaining start and end offsets as the window state will not work when there are delays. We may need to keep multiple such offsets. But this may not be a clean solution.
On Thu, Mar 5, 2015 at 2:42 PM, Milinda Pathirage <mpath...@umail.iu.edu> wrote: > Hi Yi, > > Please find my comments inline. > > On Thu, Mar 5, 2015 at 1:18 PM, Yi Pan <nickpa...@gmail.com> wrote: > >> Hi, Milinda, >> >> We have recently some discussions on the MillWheel model: >> http://www.infoq.com/presentations/millwheel. > > > Yes. Above is a very interesting talk. I asked the above question > regarding the language, just after watching the talk. I was under the > impression that we need to specify these details (handling delays) > explicitly in the query. > > >> It is very interesting talk and have one striking point that we did not >> think about before: handle late arrivals as a "correction" to the earlier >> results. Hence, if we follow that model, the late arrival problem that you >> described can be addressed in the following: >> >> a) Each window will have a closing policy: it would either be wall-clock >> based timeout, or the arrival of messages indicating that we have received >> all messages in the corresponding event time window >> > > Given that the closing policy is not explicit in the query, how we are > going to handle this. Is this policy going to be specific to a query or > system wide thing. I think I was not clear about this in the previous mail. > > >> b) Each window also keeps all the past messages it receives in the past >> windows, up to a large retention size that covers all possible late >> arrivals >> > > Are we going to keep this in local storage. Is this (keeping past > messages) really necessary in case of monotonic queries. May be you meant > to say we just keep metadata about offsets. So we can replay from Kafka (I > don't have that much experience with Kafka, but I think we can start > consuming from random offsets). > > >> c) When a window's closing policy is satisfied, the window operator always >> emits the current window results >> > > Does this means we are waiting for the window to be closed, before sending > new messages downstream? This may have performance implications, but this > will make it easy to implement the query processing. I think current > operator layer can support this style without any changes. > > >> d) When a late arrival message came, the window operator will re-emit the >> past window results to correct the previous window results >> >> > It would be better if we can do incremental updates without replaying the > whole window. But I believe there are advantages of this approach. > > >> In your example, the aggregation for the counter for window from >> 10:00-10:59 will have a "wrong" value when the window is closed by an >> arrival of message w/ 11:00 timestamp, but will be corrected later by a >> late arrival of another message in the time window from 10:00-10:59. I.e. >> if we keep all the previous window states, late arrival messages will >> simply trigger a re-computation of the aggregated counter for the window >> 10:00-10:59 and overwrite the previous result. In this model, the final >> result is always correct, as long as the late arrivals is within the large >> retention size. >> >> I have been thinking of this model and had a discussion with Julian >> yesterday. It seems that the followings are more reasonable to me: >> 1) Window operator will have a full buffered state of the stream similar >> to >> a time-varying materialized view over the retention size >> 2) Window size and termination (i.e. sliding/tumbling/hopping windows) >> will >> now determine when we emit window results (i.e. new messages/updates to >> the >> current window) to the downstream operator s.t. the operators can >> calculate >> result in time >> 3) Late arrivals will be sent to the downstream operator and triggers a >> re-computation of the past result based on the full buffered state >> >> In the above model, the window operator becomes a system feature, or an >> implementation of "StreamScan" in Calcite's term. And we do not need >> specific language support for the window semantics, with a default time >> window operator implementation that serves as a "StreamScan". All window >> definition in the query language now only dictates the semantic meaning of >> aggregation and join on top of the physical window operator which >> provides: >> a) a varying/growing materialized view; b) a driver that tells the >> aggregation/join to compute/re-compute results on-top-of the materialized >> view. >> >> >> > I will think more about this model and may have more questions about this > in future :). > > Thanks > Milinda > > >> On Wed, Mar 4, 2015 at 10:28 AM, Milinda Pathirage <mpath...@umail.iu.edu >> > >> wrote: >> >> > Hi Julian, >> > >> > I went through the draft and it covers most of our requirements. But >> > aggregation over a window will not be as simple as mentioned in the >> draft. >> > >> > In the stream extension draft we have following: >> > >> > 'How did Calcite know that the 10:00:00 sub-totals were complete at >> > > 11:00:00, so that it could emit them? It knows that rowtime is >> > increasing, >> > > and it knows that FLOOR(rowtime TO HOUR) is also increasing. So, once >> it >> > > has seen a row at or after 11:00:00, it will never see a row that will >> > > contribute to a 10:00:00 total.' >> > >> > >> > When there are delays, we can't do above. Because observing a row with >> > rowtime greater than 11:00:00 doesn't mean events from 10:00:00 to >> 10:00:59 >> > time window will not arrive after this observation. We have discussed >> this >> > in https://issues.apache.org/jira/browse/SAMZA-552. Even if we consider >> > the >> > 'system time/stream time' as mentioned in SAMZA-552, it doesn't >> guarantee >> > the absence of delays in a distributed setting. So we may need to >> > additional hints/extensions to specify extra information required to >> handle >> > complexities in window calculations. >> > >> > May be there are ways to handle this at Samza level, not in the query >> > language. >> > >> > @Chirs, @Yi >> > I got the query planner working with some dummy operators and re-writing >> > the query to add default window operators. But Julian's comments about >> > handling defaults and optimizing the query plan (moving the Delta down >> and >> > removing both Delta and Chi) got me into thinking whether enforcing CQL >> > semantics as we have in our current operator layer limits the >> flexibility >> > and increase the complexity of query plan to operator router generation. >> > Anyway, I am going to take a step back and think more about Julian's >> > comments. I'll put my thoughts into a design document for query planner. >> > >> > Thanks >> > Milinda >> > >> > >> > On Tue, Mar 3, 2015 at 3:40 PM, Julian Hyde <jul...@hydromatic.net> >> wrote: >> > >> > > Sorry to show up late to this party. I've had my head down writing a >> > > description of streaming SQL which I hoped would answer questions like >> > > this. Here is the latest draft: >> > > >> https://github.com/julianhyde/incubator-calcite/blob/chi/doc/STREAM.md >> > > >> > > I've been avoiding windows for now. They are not needed for simple >> > queries >> > > (project, filter, windowed aggregate) and I wanted to write the >> > > specification of more complex queries before I introduce them. >> > > >> > > Let's look at a simple query, filter. According to CQL, to evaluate >> > > >> > > select stream * >> > > from orders >> > > where productId = 10 (query 1) >> > > >> > > you need to convert orders to a relation over a particular window, >> apply >> > > the filter, then convert back to a stream. We could write >> > > >> > > select stream * >> > > from orders over (order by rowtime range between unbounded preceding >> > and >> > > current row) >> > > where productId = 10 (query 2) >> > > >> > > or we could write >> > > >> > > select stream * >> > > from orders over (order by rowtime range between current row and >> > current >> > > row) >> > > where productId = 10 (query 3) >> > > >> > > Very different windows, but they produce the same result, because of >> the >> > > stateless nature of Filter. So, let's suppose that the default window >> is >> > > the one I gave first, "(order by rowtime range between unbounded >> > preceding >> > > and current row)", and so query 1 is just short-hand for query 2. >> > > >> > > I currently translate query 1 to >> > > >> > > Delta >> > > Filter($1 = 10) >> > > Scan(orders) >> > > >> > > but I should really be translating to >> > > >> > > Delta >> > > Filter($1 = 10) >> > > Chi(order by $0 range between unbounded preceding and current row) >> > > Scan(orders) >> > > >> > > Delta is the "differentiation" operator and Chi is the "integration" >> > > operator. After we apply rules to push the Delta through the Filter, >> the >> > > Delta and Chi will collide and cancel each other out. >> > > >> > > Why have I not yet introduced the Chi operator? Because I have not yet >> > > dealt with a query where it makes any difference. >> > > >> > > Where it will make a difference is joins. But even for joins, I hold >> out >> > > hope that we can avoid explicit windows, most of the time. One could >> > write >> > > >> > > select stream * >> > > from orders over (order by rowtime range between current row and >> > > interval '1' hour following) >> > > join shipments >> > > on orders.orderId = shipments.orderId (query 4) >> > > >> > > but I think most people would find the following clearer: >> > > >> > > select stream * >> > > from orders >> > > join shipments >> > > on orders.orderId = shipments.orderId (query 5) >> > > and shipments.rowtime between orders.rowtime and orders.rowtime + >> > > interval '1' hour >> > > >> > > Under the covers there are still the implicit windows: >> > > >> > > select stream * >> > > from orders over (order by rowtime range between unbounded preceding >> > and >> > > current row) >> > > join shipments over (order by rowtime range between unbounded >> preceding >> > > and current row) >> > > on orders.orderId = shipments.orderId (query 6) >> > > and shipments.rowtime between orders.rowtime and orders.rowtime + >> > > interval '1' hour >> > > >> > > Query 6 is equivalent to query 5. But the system can notice the join >> > > condition involving the two streams' rowtimes and trim down the >> windows >> > > (one window to an hour, another window to just the current row) >> without >> > > changing semantics: >> > > >> > > select stream * >> > > from orders over (order by rowtime range between interval '1' hour >> > > preceding and current row) >> > > join shipments over (order by rowtime range between current row and >> > > current row) >> > > on orders.orderId = shipments.orderId (query 7) >> > > and shipments.rowtime between orders.rowtime and orders.rowtime + >> > > interval '1' hour >> > > >> > > So, my hope is that end-users will rarely need to use an explicit >> window. >> > > >> > > In the algebra, we will start introducing Chi. It will evaporate for >> > > simple queries such as Filter. It will remain for more complex queries >> > such >> > > as stream-to-stream join, because you are joining the current row of >> one >> > > stream to a time-varying relation based on the other, and Chi >> represents >> > > that "recent history of a stream" relation. >> > > >> > > Julian >> > > >> > > >> > > > On Mar 2, 2015, at 11:42 AM, Milinda Pathirage < >> mpath...@umail.iu.edu> >> > > wrote: >> > > > >> > > > Hi Yi, >> > > > >> > > > As I understand rules and re-writes basically do the same thing >> > > > (changing/re-writing the operator tree). But in case of rules this >> > > happens >> > > > during planning based on the query planner configuration. And >> > re-writing >> > > is >> > > > done on the planner output, after the query goes through the >> planner. >> > In >> > > > Calcite re-write is happening inside the interpreter and in our >> case it >> > > > will be inside the query plan to operator router conversion phase. >> > > > >> > > > Thanks >> > > > Milinda >> > > > >> > > > On Mon, Mar 2, 2015 at 2:31 PM, Yi Pan <nickpa...@gmail.com> wrote: >> > > > >> > > >> Hi, Milinda, >> > > >> >> > > >> +1 on your default window idea. One question: what's the difference >> > > between >> > > >> a rule and a re-write? >> > > >> >> > > >> Thanks! >> > > >> >> > > >> On Mon, Mar 2, 2015 at 7:14 AM, Milinda Pathirage < >> > > mpath...@umail.iu.edu> >> > > >> wrote: >> > > >> >> > > >>> @Chris >> > > >>> Yes, I was referring to that mail. Actually I was wrong about the >> > ‘Now’ >> > > >>> window, it should be a ‘Unbounded’ window for most the default >> > > scenarios >> > > >>> (Section 6.4 of >> https://cs.uwaterloo.ca/~david/cs848/stream-cql.pdf >> > ). >> > > >>> Because >> > > >>> applying a ‘Now’ window with size of 1 will double the number of >> > events >> > > >>> generated if we consider insert/delete streams. But ‘Unbounded’ >> will >> > > only >> > > >>> generate insert events. >> > > >>> >> > > >>> @Yi >> > > >>> 1. You are correct about Calcite.There is no stream-to-relation >> > > >> conversion >> > > >>> happening. But as I understand we don’t need Calcite to support >> this. >> > > We >> > > >>> can add it to our query planner as a rule or re-write. What I am >> not >> > > sure >> > > >>> is whether to use a rule or a re-write. >> > > >>> 2. There is a rule in Calcite which extract the Window out from >> the >> > > >>> Project. But I am not sure why that didn’t happen in my test. This >> > rule >> > > >> is >> > > >>> added to the planner by default. I’ll ask about this in Calcite >> > mailing >> > > >>> list. >> > > >>> >> > > >>> I think we can figure out a way to move the window to the input >> > stream >> > > if >> > > >>> Calcite can move the window out from Project. I’ll see how we can >> do >> > > >> this. >> > > >>> >> > > >>> Also I’ll go ahead and implement default windows. We can change it >> > > later >> > > >> if >> > > >>> Julian or someone from Calcite comes up with a better suggestion. >> > > >>> >> > > >>> Thanks >> > > >>> Milinda >> > > >>> >> > > >>> On Sun, Mar 1, 2015 at 8:23 PM, Yi Pan <nickpa...@gmail.com> >> wrote: >> > > >>> >> > > >>>> Hi, Milinda, >> > > >>>> >> > > >>>> Sorry to reply late on this. Here are some of my comments: >> > > >>>> 1) In Calcite's model, it seems that there is no >> stream-to-relation >> > > >>>> conversion step. In the first example where the window >> specification >> > > is >> > > >>>> missing, I like your solution to add the default LogicalNowWindow >> > > >>> operator >> > > >>>> s.t. it makes the physical operator matches the query plan. >> However, >> > > if >> > > >>>> Calcite community does not agree to add the default >> > LogicalNowWindow, >> > > >> it >> > > >>>> would be fine for us if we always insert a default "now" window >> on a >> > > >>> stream >> > > >>>> when we generate the Samza configuration. >> > > >>>> 2) I am more concerned on the other cases, where window operator >> is >> > > >> used >> > > >>> in >> > > >>>> aggregation and join. In your example of windowed aggregation in >> > > >> Calcite, >> > > >>>> window spec seems to be a decoration to the LogicalProject >> operator, >> > > >>>> instead of defining a data source to the LogicalProject >> operator. In >> > > >> the >> > > >>>> CQL model we followed, the window operator is considered as a >> query >> > > >>>> primitive that generate a data source for other relation >> operators >> > to >> > > >>>> consume. How exactly is window operator used in Calcite planner? >> > Isn't >> > > >> it >> > > >>>> much clear if the following is used? >> > > >>>> LogicalProject(EXPR$0=[CASE(>(COUNT($2), 0), >> > CAST($SUM0($2)):INTEGER, >> > > >>>> null)]) >> > > >>>> LogicalWindow(ROWS BETWEEN 2 PRECEDING AND 2 FOLLOWING) >> > > >>>> >> > > >>>> On Thu, Feb 26, 2015 at 12:18 PM, Milinda Pathirage < >> > > >>> mpath...@umail.iu.edu >> > > >>>>> >> > > >>>> wrote: >> > > >>>> >> > > >>>>> Hi devs, >> > > >>>>> >> > > >>>>> I ask about $subject in calcite-dev. You can find the archived >> > > >>> discussion >> > > >>>>> at [1]. I think your thoughts are also valuable in this >> discussion >> > in >> > > >>>>> calcite list. >> > > >>>>> >> > > >>>>> I discovered the requirement for a default window operator when >> I >> > > >> tried >> > > >>>> to >> > > >>>>> integrate streamscan (I was using tablescan prevously) into the >> > > >>> physical >> > > >>>>> plan generation logic. Because of the way we have written the >> > > >>>>> OperatorRouter API, we always need a stream-to-relation >> operator at >> > > >> the >> > > >>>>> input. But Calcite generates a query plan like following: >> > > >>>>> >> > > >>>>> LogicalDelta >> > > >>>>> LogicalProject(id=[$0], product=[$1], quantity=[$2]) >> > > >>>>> LogicalFilter(condition=[>($2, 5)]) >> > > >>>>> >> > > >>>>> StreamScan(table=[[KAFKA, ORDERS]], fields=[[0, 1, 2]]) >> > > >>>>> >> > > >>>>> If we consider LogicalFilter as a relation operator, we need >> > > >> something >> > > >>> to >> > > >>>>> convert input stream to a relation before sending the tuples >> > > >>> downstream. >> > > >>>>> In addition to this, there is a optimization where we consider >> > filter >> > > >>>>> operator as a tuple operator and have it between StreamScan and >> > > >>>>> stream-to-relation operator as a way of reducing the amount of >> > > >> messages >> > > >>>>> going downstream. >> > > >>>>> >> > > >>>>> Other scenario is windowed aggregates. Currently window spec is >> > > >>> attached >> > > >>>> to >> > > >>>>> the LogicalProject in query plan like following: >> > > >>>>> >> > > >>>>> LogicalProject(EXPR$0=[CASE(>(COUNT($2) OVER (ROWS BETWEEN 2 >> > > >> PRECEDING >> > > >>>> AND >> > > >>>>> 2 FOLLOWING), 0), CAST($SUM0($2) OVER (ROWS BETWEEN 2 PRECEDING >> > AND 2 >> > > >>>>> FOLLOWING)):INTEGER, null)]) >> > > >>>>> >> > > >>>>> I wanted to know from them whether it is possible to move window >> > > >>>> operation >> > > >>>>> just after the stream scan, so that it is compatible with our >> > > >> operator >> > > >>>>> layer. >> > > >>>>> May be there are better or easier ways to do this. So your >> comments >> > > >> are >> > > >>>>> always welcome. >> > > >>>>> >> > > >>>>> Thanks >> > > >>>>> Milinda >> > > >>>>> >> > > >>>>> >> > > >>>>> [1] >> > > >>>>> >> > > >>>>> >> > > >>>> >> > > >>> >> > > >> >> > > >> > >> http://mail-archives.apache.org/mod_mbox/incubator-calcite-dev/201502.mbox/browser >> > > >>>>> >> > > >>>>> -- >> > > >>>>> Milinda Pathirage >> > > >>>>> >> > > >>>>> PhD Student | Research Assistant >> > > >>>>> School of Informatics and Computing | Data to Insight Center >> > > >>>>> Indiana University >> > > >>>>> >> > > >>>>> twitter: milindalakmal >> > > >>>>> skype: milinda.pathirage >> > > >>>>> blog: http://milinda.pathirage.org >> > > >>>>> >> > > >>>> >> > > >>> >> > > >>> >> > > >>> >> > > >>> -- >> > > >>> Milinda Pathirage >> > > >>> >> > > >>> PhD Student | Research Assistant >> > > >>> School of Informatics and Computing | Data to Insight Center >> > > >>> Indiana University >> > > >>> >> > > >>> twitter: milindalakmal >> > > >>> skype: milinda.pathirage >> > > >>> blog: http://milinda.pathirage.org >> > > >>> >> > > >> >> > > > >> > > > >> > > > >> > > > -- >> > > > Milinda Pathirage >> > > > >> > > > PhD Student | Research Assistant >> > > > School of Informatics and Computing | Data to Insight Center >> > > > Indiana University >> > > > >> > > > twitter: milindalakmal >> > > > skype: milinda.pathirage >> > > > blog: http://milinda.pathirage.org >> > > >> > > >> > >> > >> > -- >> > Milinda Pathirage >> > >> > PhD Student | Research Assistant >> > School of Informatics and Computing | Data to Insight Center >> > Indiana University >> > >> > twitter: milindalakmal >> > skype: milinda.pathirage >> > blog: http://milinda.pathirage.org >> > >> > > > > -- > Milinda Pathirage > > PhD Student | Research Assistant > School of Informatics and Computing | Data to Insight Center > Indiana University > > twitter: milindalakmal > skype: milinda.pathirage > blog: http://milinda.pathirage.org > -- Milinda Pathirage PhD Student | Research Assistant School of Informatics and Computing | Data to Insight Center Indiana University twitter: milindalakmal skype: milinda.pathirage blog: http://milinda.pathirage.org
