This question already has answers here:
Two SQL LEFT JOINS produce incorrect result
(3 answers)
Closed 12 months ago.
I am having some troubles with a count function. The problem is given by a left join that I am not sure I am doing correctly.
Variables are:
Customer_name (buyer)
Product_code (what the customer buys)
Store (where the customer buys)
The datasets are:
Customer_df (list of customers and product codes of their purchases)
Store1_df (list of product codes per week, for Store 1)
Store2_df (list of product codes per day, for Store 2)
Final output desired:
I would like to have a table with:
col1: Customer_name;
col2: Count of items purchased in store 1;
col3: Count of items purchased in store 2;
Filters: date range
My query looks like this:
SELECT
DISTINCT
C_customer_name,
C.product_code,
COUNT(S1.product_code) AS s1_sales,
COUNT(S2.product_code) AS s2_sales,
FROM customer_df C
LEFT JOIN store1_df S1 USING(product_code)
LEFT JOIN store2_df S2 USING(product_code)
GROUP BY
customer_name, product_code
HAVING
S1_sales > 0
OR S2_sales > 0
The output I expect is something like this:
Customer_name
Product_code
Store1_weekly_sales
Store2_weekly_sales
Luigi
120012
4
8
James
100022
6
10
But instead, I get:
Customer_name
Product_code
Store1_weekly_sales
Store2_weekly_sales
Luigi
120012
290
60
James
100022
290
60
It works when instead of COUNT(product_code) I do COUNT(DSITINCT product_code) but I would like to avoid that because I would like to be able to aggregate on different timespans (e.g. if I do count distinct and take into account more than 1 week of data I will not get the right numbers)
My hypothesis are:
I am joining the tables in the wrong way
There is a problem when joining two datasets with different time aggregations
What am I doing wrong?
The reason as Philipxy indicated is common. You are getting a Cartesian result from your data thus bloating your numbers. To simplify, lets consider just a single customer purchasing one item from two stores. The first store has 3 purchases, the second store has 5 purchases. Your total count is 3 * 5. This is because for each entry in the first is also joined by the same customer id in the second. So 1st purchase is joined to second store 1-5, then second purchase joined to second store 1-5 and you can see the bloat. So, by having each store pre-query the aggregates per customer will have AT MOST, one record per customer per store (and per product as per your desired outcome).
select
c.customer_name,
AllCustProducts.Product_Code,
coalesce( PQStore1.SalesEntries, 0 ) Store1SalesEntries,
coalesce( PQStore2.SalesEntries, 0 ) Store2SalesEntries
from
customer_df c
-- now, we need all possible UNIQUE instances of
-- a given customer and product to prevent duplicates
-- for subsequent queries of sales per customer and store
JOIN
( select distinct customerid, product_code
from store1_df
union
select distinct customerid, product_code
from store2_df ) AllCustProducts
on c.customerid = AllCustProducts.customerid
-- NOW, we can join to a pre-query of sales at store 1
-- by customer id and product code. You may also want to
-- get sum( SalesDollars ) if available, just add respectively
-- to each sub-query below.
LEFT JOIN
( select
s1.customerid,
s1.product_code,
count(*) as SalesEntries
from
store1_df s1
group by
s1.customerid,
s1.product_code ) PQStore1
on AllCustProducts.customerid = PQStore1.customerid
AND AllCustProducts.product_code = PQStore1.product_code
-- now, same pre-aggregation to store 2
LEFT JOIN
( select
s2.customerid,
s2.product_code,
count(*) as SalesEntries
from
store2_df s2
group by
s2.customerid,
s2.product_code ) PQStore2
on AllCustProducts.customerid = PQStore2.customerid
AND AllCustProducts.product_code = PQStore2.product_code
No need for a group by or having since all entries in their respective pre-aggregates will result in a maximum of 1 record per unique combination. Now, as for your needs to filter by date ranges. I would just add a WHERE clause within each of the AllCustProducts, PQStore1, and PQStore2.
Yes, I know this seems simple:
SELECT DISTINCT(...)
Except, it apparently isn't
Here is my actual Query:
SELECT
DeclinationReasons.Reason,
EmployeeInformation.ID,
EmployeeInformation.Employee,
EmployeeInformation.Active,
CompletedTrainings.DecShotDate,
CompletedTrainings.DecShotLocation,
CompletedTrainings.DecReason,
CompletedTrainings.DecExplanation,
IIf([DecShotLocation]="MCS","Yes","No") AS YesMCS,
IIf([DecReason]=1,1,0) AS YesAllergy,
IIf([DecReason]=2,1,0) AS YesImmune,
IIf([DecReason]=3,1,0) AS YesAdverse,
IIf([DecReason]=4,1,0) AS YesMedical,
IIf([DecReason]=5,1,0) AS YesSpiritual,
IIf([DecReason]=6,1,0) AS YesOther,
IIf([DecReason]=7,1,0) AS YesAlready
FROM
EmployeeInformation
INNER JOIN (CompletedTrainings
LEFT JOIN DeclinationReasons ON CompletedTrainings.DecReason = DeclinationReasons.ReasonID)
ON EmployeeInformation.ID = CompletedTrainings.Employee
GROUP BY
DeclinationReasons.Reason,
EmployeeInformation.ID,
EmployeeInformation.Employee,
EmployeeInformation.Active,
CompletedTrainings.DecShotDate,
CompletedTrainings.DecShotLocation,
CompletedTrainings.DecReason,
CompletedTrainings.DecExplanation,
IIf([DecShotLocation]="MCS","Yes","No"),
IIf([DecReason]=1,1,0),
IIf([DecReason]=2,1,0),
IIf([DecReason]=3,1,0),
IIf([DecReason]=4,1,0),
IIf([DecReason]=5,1,0),
IIf([DecReason]=6,1,0),
IIf([DecReason]=7,1,0)
HAVING
((((EmployeeInformation.Active) Like -1)
AND ((CompletedTrainings.DecShotDate + 365 >= DATE())
OR (CompletedTrainings.DecShotDate IS NULL))));
This is Joining a few tables (obviously) in order to get a number of records. The problem is that if someone is duplicated on the table with a NULL in one of the date fields, and a date in another field, it pulls both the NULL and the DATE, or pulls multiple NULLS it might pull multiple dates but those are not present right at the moment.
I need the Nulls, they are actual data in this particular case, but if someone has a date and a NULL I need to pull only the newest record, I thought I could add MAX(RecordID) from the table, but that didn't change the results of the query either.
That code:
SELECT
DeclinationReasons.Reason,
EmployeeInformation.ID,
EmployeeInformation.Employee,
EmployeeInformation.Active,
MAX(CompletedTrainings.RecordID),
CompletedTrainings.DecShotDate
...
And it returned the same issue, Duplicated EmployeeInformation.ID with different DecShotDate values.
Currently it returns:
ID
Active
DecShotDate
etc. x a bunch
1
-1
date date
whatever goes
2
-1
in these
2
-1
date date
columns
These are being used in a report, that is to determine the total number of employees who fit the criteria of the report. The NULLs in DecShotDate are needed as they show people who did not refuse to get a flu vaccine in the current year, while the dates are people who did refuse.
Now I have come up with one simple solution, I could add a column to the CompletedTrainings Table that contains a date or other value, and add that to the HAVING statement. This might be the right solution as this is a yearly training questionnaire that employees have to fill out. But I am asking for advice before doing this.
Am I right in thinking I need to add a column to filter by so that older data isn't being pulled, or should I be able to do this by pulling recordID, and did I just bork that part of the query up?
Edited to add raw table views:
EmployeeInformation Table:
ID
Last
First
empID
Active
Termdate
DoH
Title
PT/FT/PD
PI
1
Doe
Jane
982
-1
date
Sr
PD
X
2
Roe
John
278
0
date
date
Jr
PD
X
3
Moe
Larry
1232
-1
date
Sr
FT
X
4
Zoe
Debbie
1424
-1
date
Sr
PT
X
DeclinationReasons Table:
ReasonID
Reason
1
Allergy
2
Already got it
3
Illness
CompletedTrainings Table:
RecordID
Employee
Training
...
DecShotdate
DecShotLocation
DecShotReason
DecExp
1
1
4
date
location
2
text
2
1
4
3
2
4
4
3
4
date
location
3
text
5
3
4
date
location
1
text
6
4
4
After some serious soul searching, I decided to use another column and filter by that.
In the end my query looks like this:
SELECT *
FROM (
(
SELECT RecordID, DecShotDate, DecShotLocation, DecReason, DecExplanation, Employee,
IIf([DecShotLocation]="MCS","Yes","No") AS YesMCS, IIf([DecReason]=1,1,0) AS YesAllergy,
IIf([DecReason]=2,1,0) AS YesImmune, IIf([DecReason]=3,1,0) AS YesAdverse,
IIf([DecReason]=4,1,0) AS YesMedical, IIf([DecReason]=5,1,0) AS YesSpiritual,
IIf([DecReason]=6,1,0) AS YesOther, IIf([DecReason]=7,1,0) AS YesAlready
FROM CompletedTrainings WHERE (CompletedDate > DATE() - 365 ) AND (Training = 69)) AS T1
LEFT JOIN
(
SELECT ID, Active FROM EmployeeInformation) AS T2 ON T1.Employee = T2.ID)
LEFT JOIN
(
SELECT Reason, ReasonID FROM DeclinationReasons) AS T3 ON T1.DecReason = T3.ReasonID;
This may not have been the best solution, but it did exactly what I needed. Which is to get the information by latest entry into the database.
Previously I had tried to use MAX(), DISTINCT(), etc. but always had a problem of multiple records being retrieved. In this case, I intentionally SELECT the most recent records first, then join them to the results of the next query, and so on. Until I have all the required data for my report.
I write this in hopes someone else finds it useful. Or even better if someone tells me why this is wrong, so as to improve my own skills.
I have two tables that I want to join together:
contracts:
id
customer_id_1
customer_id_2
customer_id_3
date
1
MAIN1
TRAN1
TRAN2
20201101
2
MAIN2
20201001
3
MAIN3
TRAN5
20200901
4
MAIN4
TRAN7
TRAN8
20200801
customers:
id
customer_id
info
date
1
MAIN1
blah
20200930
2
TRAN2
blah
20200929
3
TRAN5
blah
20200831
4
TRAN7
blah
20200801
In my contracts table, each row represents a contract with a customer, who may have 1 or more different IDs they are referred to by in the customers table. In the customers table, I have info on customers (can be zero or multiple records on different dates for each customer). I want to perform a join from contracts onto customers such that I get the most recent info available on a customer at the time a contract is recorded, ignoring any potential customer info that may be available after the contract date. I am also not interested in contracts which have no info on the customers. The main problem here is that in customers, each customer record can reference any 1 of the 3 IDs that may exist.
I currently have the following query which performs the task as intended but the problem is that is extremely slow when run on data in the 50-100k rows range. If I remove the OR statements in the INNER JOIN and just join on the the first ID, the query performs in seconds as opposed to ~ half an hour.
SELECT
DISTINCT ON (ctr.id)
ctr.id,
ctr.customer_id_1,
ctr.date AS contract_date,
cst.info,
cst.date AS info_date
FROM
contracts ctr
INNER JOIN customers cst ON (
cst.customer_id = ctr.customer_id_1
OR cst.customer_id = ctr.customer_id_2
OR cst.customer_id = ctr.customer_id_3
)
AND ctr.date >= cst.date
ORDER BY
ctr.id,
cst.date DESC
Result:
id
customer_id_1
contract_date
info
info_date
1
MAIN1
20201101
blah
20200930
3
MAIN3
20200901
blah
20200831
4
MAIN4
20200801
blah
20200801
It seems like OR statements in JOINs aren't very common (I've barely found any examples online) and I presume this is because there must be a better way of doing this. So my question is, how can this be optimised?
OR often is a performance killer in SQL predicates.
One alternative unpivots before joining:
select distinct on (ctr.id)
ctr.id,
ctr.customer_id_1,
ctr.date as contract_date,
cst.info,
cst.date as info_date
from contracts ctr
cross join lateral (values
(ctr.customer_id_1), (ctr.customer_id_2), (ctr.customer_id_3)
) as ctx(customer_id)
inner join customers cst on cst.customer_id = ctx.customer_id and ctr.date >= cst.date
order by ctr.id, cst.date desc
The use of this techniques pinpoints that your could vastly improve your data model: the relation between contracts and customers should be stored in a separate table, with each customer/contract tuple on a separate row - essentially, what the query does is virtually build that derived table in the lateral join.
I have a query is used to display information in a queue and part of that information is showing the amount of child entities (packages and labs) that belong to the parent entity (change). However instead of showing the individual counts of each type of child, they multiply with one another.
In the below case, there are supposed to be 3 labs and 18 packages, however the the multiply with one another and the output is 54 of each.
Below is the offending portion of the query.
SELECT cef.ChangeId, COUNT(pac.PackageId) AS 'Packages', COUNT(lab.LabRequestId) AS 'Labs'
FROM dbo.ChangeEvaluationForm cef
LEFT JOIN dbo.Lab
ON cef.ChangeId = Lab.ChangeId
LEFT JOIN dbo.Package pac
ON (cef.ChangeId = pac.ChangeId AND pac.PackageStatus != 6 AND pac.PackageStatus !=7)
WHERE cef.ChangeId = 255
GROUP BY cef.ChangeId
I feel like this is obvious but it's not occurring to me how to fix it so the two counts are independent of one another like to me they should be. There doesn't seem to be a scenario like this in any of my research either. Can anyone guide me in the right direction?
Because you do multiply source rows by each left join. So sometimes you have more likely cross join here.
SELECT cef.ChangeId, p.Packages, l.Labs
FROM dbo.ChangeEvaluationForm cef
OUTER APPLY(
SELECT COUNT(*) as Labs
FROM dbo.Lab
WHERE cef.ChangeId = Lab.ChangeId
) l
OUTER APPLY(
SELECT COUNT(*) AS Packages
FROM dbo.Package pac
WHERE (cef.ChangeId = pac.ChangeId AND pac.PackageStatus != 6 AND pac.PackageStatus !=7)
) p
WHERE cef.ChangeId = 255
GROUP BY cef.ChangeId
perhaps GROUP BY is not needed now.
From you question its difficult to derive what result do you expect from your query. So I presume you want following result:
+----------+----------+------+
| ChangeId | Packages | Labs |
+----------+----------+------+
| 255 | 18 | 3 |
+----------+----------+------+
Try below query if you are looking for above mentioned result.
SELECT cef.ChangeId, ISNULL(pac.PacCount, 0) AS 'Packages', ISNULL(Lab.LabCount, 0) AS 'Labs'
FROM dbo.ChangeEvaluationForm cef
LEFT JOIN (SELECT Lab.ChangeId, COUNT(*) LabCount FROM dbo.Lab GROUP BY) Lab
ON cef.ChangeId = Lab.ChangeId
LEFT JOIN (SELECT pac.ChangeId, COUNT(*) PacCount FROM dbo.Package pac WHERE pac.PackageStatus != 6 AND pac.PackageStatus !=7 GROUP BY pac.ChangeId) pac
ON cef.ChangeId = pac.ChangeId
WHERE cef.ChangeId = 255
Query Explanation:
In your query you didn't use group by, so it ended up giving you 54 as count which is Cartesian product.
In this query I tried to group by 'ChangeId' and find aggregate before joining tables. So 3 labs and 18 packages will be counted before join.
Your will also notice that I have moved PackageStatus filter before group by in pac table. So unwanted record won't mess with our count.
You start with a particular ChangeId from the dbo.ChangeEvaluationForm table (ChangeId = 255 from your example), then join to the dbo.Lab table. This join makes your result go from 1 row to 3, considering there are 3 Labs with ChangeId = 255. Your problem is on the next join, you are joining all 3 resulting rows from the previous join with the dbo.Package table, which has 18 rows for ChangeId = 255. The resulting count for columns pac.PackageId and lab.LabRequestId will then be 3 x 18 = 54.
To get what you want, there are 2 easy solutions:
Use COUNT DISTINCT instead of COUNT. This will just count the different values of pac.PackageId and lab.LabRequestId and not the repeated ones.
Split the joins into 2 subqueries and join their result (by ChangeId)
I have two columns of data I am lining up using a Full Outer Join but it includes two separate date columns which make it challenging to sort by.
Table 1 has sales rank data for a product.
Table 2 has actual sales data for the same product.
Each table may have entries for dates on which the other does not.
So envision after the full join, we end up with something like this simplified example:
ProdID L.Date P.Date Rank Units
101 null 2011-10-01 null 740
101 2011-10-02 2011-10-02 23 652
101 2011-10-03 null 32 null
Here is the query I am using to pull this data:
select L.ListID, L.ASIN, L.date, L.ranking, P.ASIN, P.POSdate, P.units from ListItem L
full outer join POSdata P on
L.ASIN = P.ASIN and
L.date = P.POSdate and
(L.ListID = 1 OR L.ASIN is null)
where (L.ASIN = 'xxxxxxxxxx' and L.ListID = 1) or
(P.ASIN = 'xxxxxxxxxx' and L.BookID is null)
order by POSdate, date
It's a bit more complex because products may appear on multiple lists so I have to account for that as well, but it returns the data I need. I am open to suggestions on improving it of course should someone have one.
The problem is, how can I sort this properly when both date columns are likely to have at least some NULLs in them. The way I am Ordering By now will not work when both columns have at one NULL.
Thanks.
ORDER BY ISNULL(P.POSdate,L.date) should do what you need I think?