This is similar to to this question which doesn't have any answers. I've read all about how to use cursors with the twitter, facebook, and disqus api's and also this article about how disqus generally built their cursors, but I still cannot seem to grok the concept of how they work and how to implement a similar solution in my own projects. Can someone explain specifically the different techniques and concepts behind them?
Lets first understand why offset pagination fails for large data sets with an example.
Clients provide two parameters limit for number of results and offset and for page offset.
For example, with offset = 40, limit = 20, we can tell the database to return the next 20 items, skipping the first 40.
Drawbacks:
Using LIMIT OFFSET doesn’t scale well for large
datasets. As the offset increases the farther you go within the
dataset, the database still has to read up to offset + count rows
from disk, before discarding the offset and only returning count
rows.
If items are being written to the dataset at a high frequency, the
page window becomes unreliable, potentially skipping or returning
duplicate results.
How Cursors solve this ?
Cursor-based pagination works by returning a pointer to a specific item in the dataset. On subsequent requests, the server returns results after the given pointer.
We will use parameters next_cursor along with limit as the parameters provided by client in this case.
Let’s assume we want to paginate from the most recent user to the oldest user.When client request for the first time , suppose we select the first page through query:
SELECT * FROM users
WHERE team_id = %team_id
ORDER BY id DESC
LIMIT %limit
Where limit is equal to limit plus one, to fetch one more result than the count specified by the client. The extra result isn’t returned in the result set, but we use the ID of the value as the next_cursor.
The response from the server would be:
{
"users": [...],
"next_cursor": "1234", # the user id of the extra result
}
The client would then provide next_cursor as cursor in the second request.
SELECT * FROM users
WHERE team_id = %team_id
AND id <= %cursor
ORDER BY id DESC
LIMIT %limit
With this, we’ve addressed the drawbacks of offset based pagination:
Instead of the window being calculated from scratch on each request based on the total number of items, we’re always fetching the next count rows after a specific reference point. If items are being written to the dataset at a high frequency, the overall position of the cursor in the set might change, but the pagination window adjusts accordingly.
This will scale well for large datasets. We’re using a WHERE clause to fetch rows with id values less than the last id from the previous page. This lets us leverage the index on the column and the database doesn’t have to read any rows that we’ve already seen.
For detailed explanation you can visit this wonderful engineering article from slack!
Here is an article about pagination: paginating-real-time-data-cursor-based-pagination
Cursors – we need to have at least one column with unique sequential values to implement cursor based pagination. This can be similar to Twitter’s max_id parameter or Facebook’s after parameter.
In general you should pass the current item or page number in the request as a param. Other usual param is the batch size of the page. Then on the server side backend you select and return the proper dataset, with an SQL query for example.
enter image description hereHere's what I am Done with. The cursor is working as a pointer and it points to that index. and limit will pick that many rows from that pointer. Let's say we have given id 10 and limit 5 then it will go to id 10 and pick 5 elements from there.
Some Graph API connections uses cursors by default. You can use 'limit' and 'before'/'after' parameters in your call. If you are still not clear, you can post your code here and I can explain with it.
Related
I'm currently building a forum alike application. Users will be able to see recent posts with the total like count. If the post is interesting to the user, they can like it as well and contribute to the total like count.
The normalized approach would be to have two tables: user_post(contains id, metadata ...), liked_post(which includes the user id + post id). When posts are getting queried, the like count would be determined with the COUNT() statement on the liked_post table grouped by the post id.
Im thinking of another approach, which requires no group by on a potential huge table. That would be to add a like_count column to the user_post table and break the normalization. This column would be always updated when a new liked_post entry gets inserted or deleted. That means: Every time a user likes a post -> there will be an update on the user_post table (increment the like_count column) + insert/delete entity in liked_post table (With a trigger or code in App layer).
Would this aggregation on the fly approach have any disadvantages, except for consistency concerns? This would enable very simple and fast select queries but Im not sure if the additional update would be an issue.
What are your thoughts ?
Im really interested in the performance impact and not if you should do this from the project begin or not.
Your idea is correct and widely used. Problem that you will face:
how do you make sure that like_count is valid? Can this number be delayed or approximated somehow?
In general you can do this following ways
update like_count within application code
update like_count by triggers
If you want to have exact values correct you could accumulate those sums by triggers or do it programatically ensuring that like count update is always within same transaction that insert to liked_posts
Using triggers it could be something like this:
CREATE FUNCTION public.update_like_count() RETURNS trigger
LANGUAGE plpgsql
AS $$
BEGIN
UPDATE user_post SET user_post.liked_count = user_post.liked_count + 1
WHERE user_post.id = NEW.post_id;
RETURN NEW;
END;
$$;
CREATE TRIGGER update_like_counts
AFTER INSERT ON public.liked_posts
FOR EACH ROW EXECUTE PROCEDURE public.update_like_count();
Also you should handle AFTER DELETE by separate trigger.
Be aware that depending on transaction isolation level you might enter concurrency problem here (if 2 inserts are done at the same time - like_count may be exactly same number for two transactions) and end up with invalid total.
So I've had a problem similar to this in the past, the solution I went with is similar to what you've described, which is having an aggregated stored value like_count. Like you mentioned the only downside would be consistency concerns however this problem exists even in the former.
The solution to something like this lies more in the application dev, so utilizing something like web-sockets to keep posts up to date, without too much fluff
When a user's browser/client loads a post they join a room with the post id, and when user interacts with a post ( like, dislike etc ) that interaction is broadcasted to all users in that room ( post id ).
Finally when it comes to finding out which users liked this post, you can query/load at the point of when the user clicks to find out. ~ cheers
Scenario:
I am displaying a table of records. It initially displays the first 500 with "show more" at the bottom, which returns the next 500.
Issue:
If between initial display and clicking "show more" 1 record is added, that will cause "order by date, offset 500, limit 500" to overlap by 1 row.
I'd like to "order by date, offset until 'id of last row shown', limit 500"
My row IDs are UUIDs. I am open to alternative approaches that achieve the same result.
If you can order by ID, you can paginate using
where id > $last_seen_id limit 500
but that's not going to be useful where you're sorting by date.
Sort stability!
I really hope that "date" actually means "timestamp" though, otherwise your ordering will be unstable and you can miss rows in pagination; you'll have to order by date, id to get stable ordering if it's really a date, and should probably do so even for timestamp.
State on client
One option is to push the state out to the client. Have the client remember the last-seen (date,id) tuple, and use:
where date > $last_seen_date and id > $last_seen_id limit 500
Cursors
Do you care about scalability? If not, you can use a server-side cursor. Declare the cursor for the full query, without the LIMIT. Then FETCH chunks of rows as requested. To do this your app must have a way to consistently bind a connection to a specific user's requests, though, and not to reset that connection or return it to the pool between requests. This might not be practical with your pool/framework, but is probably the best solution if you can do it.
Temp tables
Another even less scalable option is to CREATE TABLE sessiondata.myuser_myrequest_blah AS SELECT .... then paginate that table. It's guaranteed not to change. This avoids the difficulty of needing to keep a consistent connection across requests, but will have a very slow first-request response time and is completely impractical for large user counts or large amounts of data.
Related questions
Handling paging with changing sort orders
Using "Cursors" for paging in PostgreSQL
How to provide an API client with 1,000,000 database results?
i think you can use a subquery in the where to accomplish this.
e.g. given you're paginating through a users table, and you want the records after a given user:
SELECT *
FROM users
WHERE created_at > (
SELECT created_at
FROM users
WHERE users.id = '00000000-1111-2222-3333-444444444444'
LIMIT 1
)
ORDER BY created_at DESC limit 5;
Here is my requirement.
Front (Client) end will do a search based on predefined conditions (for instance: customer id, account number, first name, last name, etc). I need to get the data corresponding to this request from a db2 database and send it back to them (Server). We use CICS channels and containers to pass requests and responses between the Client and Server.
Front end needs the data ordered by: Receive date descending, Customer id Ascending, Account number Ascending. Data are fetched in pages of 500 records. For example, if for a search request from front end would retrieve 50,000 records from the db2 database, we need to return this data in 500 record "pages". For pagination concept, we use the field security deposit number which is primary key to our database but the sorting order is not based on this field.
I would like to know whether we can use scrollable cursor logic in CICS to implement pagination.
Please note that I do not prefer to go for internal array bubble sort to send the data in response as it would degrade performance. I like to do it via query logic.any thoughts?
Example (Initial Front end input request):
Customer id : A
First time request (To identify whether it is first time or next or previous request for pagination)
First security deposit number : 0
Last security deposit number : 0
Since this is first time request, both this field will be having zero from front end and we need to retrieve records from database based on condition of security deposit > 0
Db2 database:
There are 700 records for this criteria
Mainframe response for first time:
We will send the first 500 records
Front end will then send request for getting next set of records which will contain:
Customer id: A
Next request
First security deposit numbr: 0
Last security deposit number : 17980
So for this detail, if I query my datbase based on security deposit number > 17980, it may result in duplicate records listing in the screen once again since our sorting order in database is not based on security deposit number
How to impelement this logic??
Many Client/Server applications in an IBM Mainframe environment involve psuedo conversational CICS transactions.
If you are using CICS in psueudo conversational mode it
is not possible for the Server to hold cusors when it RETURNs to the Client. Therefore scrollable cusors
are of little use in this environment. So to answer your basic question: No scrollable cursors cannot be used here.
The "trick" here is to create an SQL predicate in the Server that is restartable. It will then pick up rows in the correct order from any given
stating point. When the Client calls your Server it must pass all of the positioning information to your Server.
Typically, on a first call from a Client all of the positioning values are set to cause the cursor to
position itself starting with what must be the the first row. The Server then pulls in a "page" worth of data
and returns it to the Client. On the next page forward request the Client sets these positioning values to
the last row it displayed and calls the Server for the next "page" of data.
In your situation I would assume that the page forward cursor would look something like this, all the
variables prefixed with RESTART... are what the Client must provide to the Server to start the cursor
in the correct position.
DECLARE CURSOR Page-forward FOR
SELECT Receive_Date, Customer_id, Account_Nbr, Security_Dep_Id
FROM Table_Name
WHERE ( (Receive_Date < :RESTART-RCV-DT)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id > :RESTART-CUSTOMER-ID)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id = :RESTART-CUSTOMER-ID AND
Account_Nbr > :RESTART-ACCT-NBR)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id = :RESTART-CUSTOMER-ID AND
Account_Nbr = :RESTART-ACCT-NBR AND
Security_Dep_Id > :RESTART-SEC-DEP-ID))
ORDER BY 1 DESC, 2 ASC , 3 ASC, 4 ASC
For the initial call the Client would have passed something like '9999-12-31' as the RESTART-RCV-DT, zero
for the RESTART-CUSTOMER-ID, RESTART-ACCT-NBR and SEC-DEP-ID (assuming these are all numeric). If you look at
the cursor predicate carefully you can verify that there cannot be any rows prior to these values - therefore this
will return the first page of data. If the Client needs to page forward after this, it must tell the Server to start
with the next row after the last one it received. To do this it would populate the RESTART... variables with
the values from the last row on the page it just
displayed. This process will drive the cursor selects forward one page at a time.
When paging up, the process is reversed (you will need a second cursor to support this, and the Client needs to tell you which direction to page: Forward or Back). The Client
will need to populate the RESTART variables with the first row it recieved from the Server. The trick
for the Server on a page up request is to return the data
to the Client in reverse order. You may have to populate the data page passed back
to the Client in reverse order (ie. put the first row retrieved into the last row of the paging area shared between
the Client and the Server). The page backward cursor would look something like:
DECLARE CURSOR Page-backward FOR
SELECT Receive_Date, Customer_id, Account_Nbr, Security_Dep_Id
FROM Table_Name
WHERE ( (Receive_Date > :RESTART-RCV-DT)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id < :RESTART-CUSTOMER-ID)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id = :RESTART-CUSTOMER-ID AND
Account_Nbr < :RESTART-ACCT-NBR)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id = :RESTART-CUSTOMER-ID AND
Account_Nbr = :RESTART-ACCT-NBR AND
Security_Dep_Id < :RESTART-SEC-DEP-ID))
ORDER BY 1 ASC, 2 DESC , 3 DESC, 4 DESC
As has been pointed out in other answers, this type of paging process does not manage or detect concurrent
updates to the database that may occur duing paging transactions. That is another topic for another day...
Developing Restartable Cursors
The the key to building a paging Server is to develop a cursor that is restartable from a set of values received
from a Client transaction. This leaves control of cursor positioning and direction with the Client.
It also means the Client must receive all critical positioning data from the Server even though
the Client might not actually
use these data for any other purpose (e.g. From your question I got the impression that the Client may not require
the Security Deposit Id except to supply as a positioning parameter for your Server)
To build a paging Server you need to know
what the required sorting order of the data are (e.g. Receive Date Descending then Customer Id Ascending then
Account Number Ascending).
You also need know the set of data that uniquely identify a row
returned by the cursor. In your case that would be the Security Deposit Id (this is the primary key for the
table you are selecting from so it must be unique for each and every row in that table). Knowing this you then build a
cursor predicate (the stuff in the WHERE clause) that will return data needed by the Client in the required sort order that
also includes
the full positioning key (i.e. Security Deposit Id). In the event that two or more returned rows may contain identical data if
the final positioning key were elimiminated makes it important that the positioning key be included as a sort condition.
It doesn't matter if it is ascending or descending, but it needs to be included on the sort to ensure consistent
order of data retrieval.
A fairly simple formula may be followed to build the predicate for a restartable cusor needed to
support paging Servers. Basically this is a cascade of "OR" clauses connecting a series of "AND" clauses
that become progressively more selective following the sort order required by the Client and end up with the positioning
key.
To see how this works consider how the query for your Server might be developed...
Start with the column from the sort order that changes least often...
SELECT ...
FROM ...
WHERE Receive_Date < restart value
This will retrieve all rows prior to the specified restart Receieve date regardless of what the other
column restart values are (e.g. Customer ID's can range from minimum to maximum values, as long as the Receive Date
is less than any Receive Date "seen" so far). Since this column only changes value after all subortinate sort columns values
have been exausted you can be sure that this does not pick up any rows prior to the full restart key.
But what about those rows that occur on the same date as the restart request but have a
larger Customer Id? These can be picked up with....
SELECT ...
FROM ...
WHERE Receive_Date = restart value AND
Customer_id > restart value
What about those where the Receive Date and Customer Id are the same as the restart key but have
a larger Account Number? These can be picked up with...
SELECT ...
FROM ...
WHERE Receive_Date = restart value AND
Customer_Id = restart value AND
Account_Nbr > restart value
Continue this pattern until the full restart key has been processed. Notice that the inequality
signs are determined by the sort order. Use < when the column is sorted Descending and > when Ascending.
Also notice that the SELECT and FROM clauses
are exactly the same for each query - which means you can put them all together using OR conjuctions...
SELECT Receive_Date, Customer_id, Account_Nbr, Security_Dep_Id
FROM Table_Name
WHERE ( (Receive_Date < :RESTART-RCV-DT)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id > :RESTART-CUSTOMER-ID)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id = :RESTART-CUSTOMER-ID AND
Account_Nbr > :RESTART-ACCT-NBR)
OR (Receive_Date = :RESTART-RCV-DT AND
Customer_Id = :RESTART-CUSTOMER-ID AND
Account_Nbr = :RESTART-ACCT-NBR AND
Security_Dep_Id > :RESTART-SEC-DEP-ID))
ORDER BY 1 DESC, 2 ASC , 3 ASC, 4 ASC
There you go... a restartable cursor for forward paging. Construction of the cursor for backward paging follows a similar pattern, just flip the
sort orders and repeat.
A simplistic approach: Write your SQL to retrieve data according to your criteria, in the sort order you specify. Then only retrieve the keys to the rows you want. Save the keys somewhere you will have access to upon subsequent invocations of your transaction. Look into multi-row select in DB2. Also understand pseudo-conversational programming techniques in CICS.
And now we get to the design implications Bill Woodger mentions, that you do not specify in your question, and which are the reason I'm just hitting the high points of a simplistic approach.
If changes to your result set occur between one invocation and the next, your results will not reflect those changes. You must decide if this is important.
You mention a "front end" but do not specify what it is. If it is a BMS application, you may be able to save the keys in your commarea or in a container. If your front end is a distributed application invoking your transactions via CICS Web Services or CICS Web Support or MQ or raw sockets or whatever, you must design a mechanism to store those keys such that you can uniquely retrieve them — perhaps by sending a contrived key back to the distributed application which it must supply upon subsequent invocations. Then you must have some process to clean up your key store.
Creating a solution to your problem that is unique in your IT shop is not something to be done in isolation. You must involve others who will be tasked with maintaining your application, there may be a group external to your project tasked with making such decisions, there may be infrastructure issues with your solution.
So this isn't so much as an answer to your question as it is an elaboration upon why you may not get an answer, or at least the answer you seem to desire.
I need to setup worldship to pull from one of our postgres databases. I need to have it so that the packages are sorted by id. I have no way (that i am aware of) of having worldship send the order by clause so I need to have the default for the records returned to be returned by id.
On a second note I have no idea how postgres default sorts it looks like it by the last time the record was changed so if i write a two records id 1,2 then change record 2 when I run the query it returns them with record 2 being first.
Rows are returned in an unspecified order, per sql specs, unless you add an order by clause. In Postgres, that means you'll get rows in, basically, the order that live rows read on the disk.
If you want a consistent order without needing to add an order by clause, create a view as suggested in Jack's comment.
There is no such thing as a "default sort". Rows in a table are not sorted.
You could fake this with a view (as suggested by Jack Maney) there is no way you can influence the order of the rows that are returned.
But if you do that, be aware that adding an additional ORDER BY to a SELECT based on that view will sort the data twice.
Another option might be to run the CLUSTER command on that table to physically order the rows on the disk according to the column you want. But this sill does not guarantee that the rows are returned in that order. Not even with a plain SELECT * FROM your_table (but chances are reasonably high for that).
You will need to re-run this statement on a regular basis because the order created by the CLUSTER command is not automatically maintained.
For what it's worth, which probably isn't much, from my testing, it appears that PostgreSQL's "default" ordering is based on the time the records were last updated. The most recently updated records will appear last. Note that I couldn't find any documentation to support this. It's just what I've found from my own testing.
You could eventually use a sorted index, which should guarantee you order of retrieved rows in case the query plan hits the index, or if you force it, but this approach will be more than circuitous :). ORDER BY clause is the way to go as mentioned already.
I started using Zend_Paginator,
it works everything fine but I noticed that there is one more query which slows the load time down.
The additional query:
SELECT COUNT(1) AS `zend_paginator_row_count` FROM `content`
The normal query:
SELECT `content`.`id`, `content`.`name` FROM `content` LIMIT 2
PHP:
$adapter = new Zend_Paginator_Adapter_DbSelect($table->select()->from($table, array('id', 'name')));
$paginator = new Zend_Paginator($adapter);
Could I merge the two querys into one (for better performance)?
Make sure you have an index on one or more int-based columns of the selected table. So the count query will not have much of a performance inpact. You can use setRowCount() to provide the count (if you have it).
From http://framework.zend.com/manual/en/zend.paginator.usage.html :
Note: Instead of selecting every
matching row of a given query, the
DbSelect and DbTableSelect adapters
retrieve only the smallest amount of
data necessary for displaying the
current page. Because of this, a
second query is dynamically generated
to determine the total number of
matching rows. However, it is possible
to directly supply a count or count
query yourself. See the setRowCount()
method in the DbSelect adapter for
more information.
Merging the two wouldn't really have much of a performance increase if any. The actual select content has the limit statement in it (as you are trying to get a subset of the entire table in the database) where the count needs to count all rows in the database. The reason it is done like this is to prevent having to select a very large set of data simply to get the count.