How about trying...
..... avg(for $irrc in $i.RR_clipped return $irrc) .....
That could make the compiler happier, potentially.
(Right now it isn't "feeling the love" in terms of the type of avg's
argument - not sure why count worked.)
Cheers,
Mike
On 2/22/16 2:03 PM, Ian Maxon wrote:
Regarding the latest query about computing some summary statistics from the
min-maxed data, Yiran and I just finished meeting and we were able to come
up with a work around.
The query was as follows:
use dataverse mt16
for $i in dataset HRM_binned_clipped return {
"row_id": $i.row_id,
"sid": $i.sid,
"gdate": $i.gdate,
"gday": $i.gday,
"timebin": $i.timebin,
"stdv_RR_clipped": avg($i.RR_clipped)
}
However, it would fail like this:
Type of argument in function-call: asterix:avg, Args:[function-call:
asterix:field-access-by-name, Args:[%0->$$0, AString: {RR_clipped}]] should
be a collection type instead of ANY [AlgebricksException]
Because HRM_binned_clipped was open and there is some sort of bug in the
avg() function for this. count() works fine.
The work around, is to just copy everything into a closed dataset, where
RR_clipped is of type [double]. The query then works.
Again though this is kind of a work-around, to a work around. The original
query:
declare function minmax($x){
let $stdv := (avg(for $z in $x return $z*$z) - avg($x) * avg($x))^(0.5)
for $y in $x
where $y < (2*$stdv) + avg($x)
and $y > avg($x) - (2*$stdv)
return $y
}
for $i in dataset HRM
group by $sid := $i.sid, $gdate := $i.date, $gday := $i.day, $timebin :=
interval-bin($i.time, time("00:00:00"), day-time-duration("PT1M")) with $i
return {
"sid": $sid,
"gdate": $gdate,
"gday": $gday,
"timebin": $timebin,
"stdv": (avg(for $ii in minmax(for $jj in $i return $jj.RR) return $ii.RR *
$ii.RR) - avg(for $ii in minmax(for $jj in $i return $jj.RR) return $ii.RR)
* avg(for $ii in minmax(for $jj in $i return $jj.RR) return $ii.RR))^(0.5)
}
Expresses the same thing without any intermediate datasets. This query
fails in compilation (only with the avgs added). I need to get the stack
from the version that Yiran is running on though, I can't reproduce it on
master; it fails in a different way.
-Ian
On Sun, Feb 21, 2016 at 2:55 PM, Mike Carey <[email protected]> wrote:
ARGH!!!! This is what we would like you to *not* have to do. Sorry...
Our aim is to be the Big Data antidote....
@Yingyi: Maybe you could take a quick peek at the query issue and
see if there is any low-hanging hope there?
@Yiran: How big are your windows, typically? (Number of data points.)
Cheers,
Mike
On 2/21/16 2:42 PM, Yiran Wang wrote:
Thank you Mike for your update and suggestions! And thank you Ian again
for working with me.
A little update from my end:
I have been working on query (1) over the weekend for a work-around. I
tried to simultaneously calculate the stdev on the new copy of the list of
values with outliers removed, which in nature is the same as the query (2).
So I ran into the same problem that the query did not compile.
What I did was to export the entire dataset with the outliers removed
into Excel and calculate the stdev in Excel. However, the entire dataset is
now 363,466 x 200+ in dimension. Though they do not exceed the row x col
limit in Excel, the memory of my computer is not big enough to do anything
useful without crashing. So I've been breaking the dataset into smaller
parts and working on each separately.
Yiran
On Sun, Feb 21, 2016 at 2:10 PM, Mike Carey <[email protected] <mailto:
[email protected]>> wrote:
Ian,
Thanks working with Yiran on this! I think there is "good" and
bad news w.r.t these queries:
- The bad news is that they go beyond what we are likely to
optimize at all well at present,
as they go beyond what typical DB aggregate functions like
min/max/avg/count/sum do.
(I would try forming the groups and then doing these things on
the groups, but saying
them in AQL will be tricky, and may lead to queries that hit
edge cases in the optimizer.
For some of these my thought was to try using a positional
variable within a group...?)
- The "good" news (only for AsterixDB) is that this is exactly
the sort of inspiration that we
are looking for in terms of understanding how to better for
query-based analytics in real
use cases (and this is a very real one!).
To quote a short paper I reviewed just this AM on SQL queries kind
of like these: "Percentage
queries are more complex than their conventional counterparts and
introduce new challenges
for optimization." (The paper didn't have an applicable solution
for us, sadly.)
A more general facility that I wish we could offer was to do
grouping in AsterixDB but then
have the ability to pass a group to (e.g.) R and then get results
back for the group. When
groups are small-ish (like Yiran's windows) that would be pretty
cool - then one could do
all sorts of advanced things per group.
Cheers,
Mike
On 2/21/16 12:35 AM, Ian Maxon wrote:
Yiran and I came up with possible answers for these...
For 1) , a function could be used that looks something like this:
declare function minmax($x){
let $stdv := (avg(for $z in $x return $z*$z) - avg($x) *
avg($x))^(0.5)
for $y in $x
where $y < (2*$stdv) + avg($x)
and $y > avg($x) - (2*$stdv)
return $y
}
And then applied to return a new copy of the list of values,
removing ones
that are outside of 2 stdev.
For 2), we also did come up with a potential solution ,but the query
fails
to compile (Filed ashttps://
issues.apache.org/jira/browse/ASTERIXDB-1308 )
Any thoughts on these queries would be welcome :) 1) especially seems
inefficient to do as a function.
- Ian
On Fri, Feb 19, 2016 at 3:37 PM, Yiran Wang<[email protected]>
<mailto:[email protected]> wrote:
Hi Asterix team,
I have two queries I'm struggling with. I'm hoping you could
provide a
direction for me. Thanks in advance!
Here is what the data structure looks like:
create type HRMType as closed {
row_id: int32,
sid: int32,
date: date,
day: int32,
time: time,
bpm: int32,
RR: float
};
create dataset HRM (HRMType)
primary key row_id;
Previously I have used the time bin function to calculate the
standard
deviation of bpm for each time bin:
for $i in dataset HRM
group by $sid := $i.sid, $gdate := $i.date, $gday := $i.day,
$timebin :=
interval-bin($i.time, time("00:00:00"), day-time-duration("PT1M"))
with $i
return {
"sid": $sid,
"gdate": $gdate,
"gday": $gday,
"timebin": $timebin,
"stdv": (avg(for $ii in $i return $ii.RR * $ii.RR) - avg(for $ii in
$i
return $ii.RR) * avg(for $ii in $i return $ii.RR))^(0.5)};
Now I have two things I am hoping to do but need help with:
1. For each 1-min time bin, remove the bpm values that are above
the top
5% or below the bottom 5%. I thought about using the min/max
function for a
few times to achieve this, but realized that it was not a good idea
because
in each time bin, the number of instances are not always the same.
So for
each 1-min time bin, we do need to calculate the 5% and 95%
threshold, and
remove instances accordingly, which I don't know how to do.
2. After removing the outliers of bpm for each 1-min time bin,
calculate a
median absolute deviation (MAD) for each 1-min time bin (as another
measure
of variation besides the standard deviation). MAD =
median(abs(x-median(x)). I'm not sure how to write a query to do
the median
function in Asterix.
Thank you so much in advance. Let me know if my questions are clear.
Yiran
--
Best,
Yiran
--
You received this message because you are subscribed to the Google
Groups
"asterixdb-dev" group.
To unsubscribe from this group and stop receiving emails from it,
send an
email [email protected]
<mailto:[email protected]>.
For more options, visithttps://groups.google.com/d/optout.
--
Best,
Yiran
