I'm writing a interface to query pagination data from Hbase table ,I query pagination data by some conditions, but it's very slow .My rowkey like this : 12345678:yyyy-mm-dd , length of 8 random Numbers and date .I try to use Redis cache all rowkeys and do pagination in it , but it's difficult to query data by the other conditions .
I also consider to design the secondary index in Hbase , and I discuss it with colleagues ,they think the secondary index is hard to maintain .
So , who can give me some ideas?
First thing, AFAIK random number + date pattern of rowkey may lead to hotspotting, if you scale with large data.
Regarding Pagination :
I'd offer solr + hbase if you are using cloudera then its cloudera search. It gives good performance(proved in our case) while querying 100 per page and with webservice call we have populated angularjs dashboard.
Also, most important thing is you can move back and forth between pages with out any issues..
Below diagram describes that.
To achieve this, you need to create collections(from hbase data) and can use solrj api
Hbase alone with scan api doesn't work for quick queries.
Apart from that, Please see my answer. Which is more insightful with implementation details...
How to achieve pagination in HBase?
Hbase only solution could be Hindex (co-processor based solution)
Link explains more in detail
Hindex architecture :
In Hbase to achieve good read performance you want your data retrieved by small number of gets (requests for single row) or a small scan (request over range of rows). Hbase stores your data sorted by key, so most important idea is to come up with such row key that would allow it.
Your key seems to contain only random integer and date so I assume that your queries are about pagination over records marked with time.
First idea is that in typical pagination scenario you access just 1 page at a time and navigate from page 1 to page 2 to page 3 etc. Given you want to paginate over all records for date 2015-08-16 you could use a scan of 50 rows with start key '\0:2015-08-16' (as it is smaller than any row in 2015-08-16) to retrieve first page. After retrieval of first page you have last key from a first page, say '12345:2015-08-16'. You can use it (or 12346:2015-08-16) to make another scan with start key 12346:2015-08-16 of 50 rows to retrieve page 2 and so on. So using this approach you query your pages fast as a single scan with predefined number of returned rows. So you can use last page row key as a parameter to paging API or just put last row key in redis so next paging API call will find it there.
All this works perfectly well until some user comes in and clicks directly to page 100. Or try to click on page 5 when he was on page 2. In such scenario you can use similar scan with nSkippedPages * 50 rows. This will not be as fast as a sequential access, but it's not a usual page usage pattern. You can use redis then to cache last row of the page result in a structure like pageNumber -> rowKey. Then if next user comes and clicks on page 100, it will see same performance as is in usual click page 1- click page 2- click page 3 scenario.
Then to make things more fast for users which click on page 99 first time, you could write a separate daemon which retrieves every 50th row and puts result in redis as a page index. Then launch it every 10-15 minutes and say that your page index has at most 10-15 minutes stale data.
You also can design a separate API which preloads row keys for a bulk of N pages (say about 100 pages, it could be async e.g. don't wait for actual preload to complete). What it will do is just a scan with KeyOnlyFilter and 50*N results and then selection of rowkeys for each page. So it accepts rowkey and populates redis with rowkey cache for N pages. Then when user walks in on a first page you fetch first 100 pages row keys for him so when he clicks on some page link seen on page, page start row key will be available. With right bulk size of preload you could approach your required latency.
Limit could be implemented using Scan.setMaxResults() or using PageFilter.
"skip nPages * 50 rows" and especially "output every 50th row" functionality seems to be trickier e.g. for latter you may end-up performing full scan which retrieves the keys or writing map-reduce to do it and for first it is not clear how to do it without sending rows over network since request can be distributed across several regions.
If you are looking for secondary indexes that are maintained in HBase there are several open source options (Splice Machine, Lilly, etc.). You can do index lookups in a few milliseconds.
Related
I have an application which calls the database multiple times to achieve one simple goal.
A little information about this application; In short, the application scrapes data from a webpage & stores specific information from this page into a database. The important information in this query is: Player name, Position. There can be multiple sitting at one specific position, kill points & Class
Player name has every potential to change or remain the same every day
Regarding the Position, there can be multiple sitting in one position
Kill points has the potential to increase or remain the same every day
Class, there is only 2 possibilities that a name can be, Ex: A can change to B or remain A (same in reverse), but cannot be C,D,E,F
The player name can change at any particular day, Position can also change dependent on the kill point increase from the last update which spins back around to the goal. This is to search the database day by day, from the current date to as far back as 2021-02-22 starting at the most recent entry for a player name and back track to the previous day to check if that player name is still the same or has changed.
What is being used as a main reference to the change is the kill points. As the days go on, this number will either be the exact same or increase, it can never decrease.
So now onto the implementation of this application.
The first query which runs finds the most recent entry for the player name
SELECT TOP(1) * FROM [changes] WHERE [CharacterName]=#charname AND [Territory]=#territory AND [Archived]=0 ORDER BY [Recorded] DESC
Then continue to check the previous days entries with the following query:
SELECT TOP(1) * FROM [changes] WHERE [Territory]=#territory AND [CharacterName]=#charname AND [Recorded]=#searchdate AND ([Class] LIKE '%{Class}%' OR [Class] LIKE '%{GetOpposite(Class)}%' AND [Archived]=0 )
If no results are found, will then proceed to find an alternative name with the following query:
SELECT TOP(5) * FROM [changes] WHERE [Kills] <= #kills AND [Recorded]='{Data.Recorded.AddDays(-1):yyyy-MM-dd}' AND [Territory]=#territory AND [Mode]=#mode AND ([Class] LIKE #original OR [Class] LIKE #opposite) AND [Archived]=0 ORDER BY [Kills] DESC
The aim of the query above is to get the top 5 entries that are the closest possible matches & Then cross references with the day ahead
SELECT COUNT(*) FROM [changes] WHERE [CharacterName]=#CharacterName AND [Territory]=#Territory AND [Recorded]=#SearchedDate AND [Archived]=0
So with checking the day ahead, if the character name is not found in the day ahead, then this is considered to be the old player name for this specific character, else after searching all 5 of the results and they are all found to be present in the day aheads searches, then this name is considered to be new to the table.
Now with the date this application started to run up to today's date which is over 400 individual queries on the database to achieve one goal.
It is also worth a noting that this table grows by 14,400 - 14,500 Rows each and every day.
The overall question to this specific? Is it possible to bring all these queries into less calls onto the database, reduce queries & improve performance?
What you can do to improve performance will be based on what parts of the application stack you can manipulate. Things to try:
Store Less Data - Database content retrieval speed is largely based on how well the database is ordered/normalized and just how much data needs to be searched for each query. Managing a cache of prior scraped pages and only storing data when there's been a change between the current scrape and the last one would guarantee less redundant requests to the db.
Separate specific classes of data - Separating data into dedicated tables would allow you to query a specific table for a specific character, etc... effectively removing one where clause.
Reduce time between queries - Less incoming concurrent requests means less resource contention and faster response times to prior requests.
Use another data structure - The only reason you're using top() is because you need data ordered in some specific way (most-recent, etc...). If you just used a code data structure that keeps the data ordered and still easily-query-able you could then perhaps offload some sql requests to this structure instead of the db.
The suggestions above are not exhaustive, but what you do to improve performance is largely a function of what in the application stack you have the ability to modify.
I'm have a DynamoDB table called URLArray that contains a list of URL's (myURL) and a unique video name (myKey).
I need to do two things:
When a user clicks the next video button, a random entry needs to be selected from this URLArray. There could be potentially tens of thousands of rows.
The user is logged into the app. Everytime they finish watching a video, the video's unique video name is recorded. So....when the user has seen a video, its added to a list in a table called Users under the user's info row.
Soo...This random entry that gets selected when the user clicks the next video button in point 1, has to be compared to the list of videos they've already seen. To make sure that it doesn't randomly appear again for that particular user.
I do something woefully inefficient so far, that works, but it's not great:
By the way i'm using AppSync + GraphQL to interact with the DynamoDB table. I first get a local copy of the URLArray:
//Gets a list of the Key/URL pairs in the UrlArrays table in GraphQL ****IN CONSTRUCTOR, so we have this URLArray data when componentDidMount()****
listUrlArrays = async () => {
try {
URLData = await API.graphql(graphqlOperation(ListUrlArrays)); //GraphQL query
//URLData[] is available in the entire class
this.setState({urlArrayLength: apiData.data.listURLArrays.items.length}); //gets the length of URLArray (i.e. how many videos are in the database)
}
}
As an overview, when user clicks for the next video:
//When clicking next video
async nextVideo(){
await this.logVideosSeen(); //add myKey to the list of videos in *Users* table the logged in user has now seen
await this.getURL(); //get the NEXT upcoming video's details, for Video Player to play and make sure it's not been seen before
}
//This will update the 'listOfVideosSeen[]' in Users table with videos unique myKey, the logged in user has seen
logVideosSeen = async () => {
.......
}
async getURL() {
var dbIndex = this.getUniqueRandomNumber(this.state.urlArrayLength); //Choose a number between 0 and N number of videos in URLArray
//the hasVideoBeenSeen() basically gets the list of videos a user has already seen from `Users` table with the GraphQL getusers command, and creates a local copy of this list (can get big). I use javascripts indexOf() to check whether myKey already exists in the list
while(await this.hasVideoBeenSeen(this.state.URLData[dbIndex].myKey)) //while true i.e. user has seen that video before
{
dbIndex = this.getUniqueRandomNumber(this.state.urlArrayLength); //get another random number to fetch a new myKey
}
//If false, we'll exit the loop and know we've got a not seen before myKey, proceed to set to play...
if(dbIndex != null){
this.setState({ playURL: this.state.URLData[dbIndex].vidURL }); //Retrieve the URL from the local URLArray that we're going to play (i.e. the next video to come)
}
}
I can share a little more code if needed, but essentially I wanted to know how to:
Let a Lambda function select a random number based on the current URLArray size (i may need to keep a local copy of URLArray anyway). But i think point 2 here is where it's really inefficient..
Let a Lambda function check (the while loop) against the Users table whether myKey has already been seen. Mainly to shift this computational burden to the cloud instead of the local device the app runs on.
AFTER A THINK................
Thanks for the suggestion Seth. I have been thinking about it for some time, and while the randomness requirement still holds true, I think there is some truth in what you’ve suggested. The reason I need randomness, is so that 2 users sat side by side for example, can’t predict which video is coming next. It shouldn’t be a predictable sequence of videos. I'm not sure I can use Scan function with AWS Amplify/GraphQL. So remember there’s 2 things going on here: (1) a video upload, recording it in the URLArray sensibly for future reference. (2) users viewing a previously unseen random video and then moving onto another unseen random video
*(1)
I like your idea of using a number to index the URLArray, and it’s helped to make life a bit easier. So the first URL being at index 0, the next at 1 etc…
My thinking here (to avoid me doing a ListUrlArrays() and bringing the WHOLE array locally to the phone), is to create a GSI called VideoNumber for the URLArray table. This will be the unique VideoNumber column with a number 0-N. So imagine the diagram above having another column called VideoNumber. Row 1 having VideoNumber set to 0, Row 2 having VideoNumber set to 1 etc… THEN all I would need to do, is locally on the device, generate a random number between 0-N, call a getURLArrayIdbyVideoNumber() query specific for that GSI, with the number that we just generated, and it’ll unlock the information I need from the row. Voila! I think that shifts most of that heavy burden away now.
Question: Before each video is uploaded, how do I easily get the current total number of rows N in the table (or row count)? I would then increment it by one.
The other thing I can do is save this current count number in another DynamoDB table that I use for persisted data, read the number from there before upload, and write an N+1 after upload to increment it (2 DynamoDB operations per upload). It’s not ideal.
*(2)
When a user has finished watching a video, I can log in a list (under the users information in DynamoDB), which video’s they’ve already seen. So for example this could now be a seen list: [3,12,73,108,57] for the 5 videos they’ve seen so far. When the user clicks nextVideo() we’ll generate a random newNumber, and straight away compare that with any number in the seen list. I use seenlist.indexOf(newNumber) and it will, either go again or stop if the newNumber doesn’t exist in the list. THEN I can go through the GSI query, and retrieve the relevant information to display the video from URLArray.
I think that this indexOf() is the biggest computational burden on the device, and obviously gets a little slower as the seenList increases. But it should be quicker with pure integer numbers then an alphanumeric myKey as I was using before. Any other suggestion would be welcome :)
I’ve yet to try it, but it was just an idea, as I need to keep the random element. But first, do you know how I can easily find the number of rows or table count of URLArray?
I think you'll have an easier time coming up with a solution to this problem if you drop the randomness requirement. It sounds like the more important requirement is presenting the user with a video they haven't seen before.
If that's correct, it sounds like your access pattern could be stated as
Fetch previously unseen video for user
which is an easier problem to solve.
Unlike SQL databases, there are often many ways to implement a given access pattern in DynamoDB. My answer here is just one way.
Imagine your URLArray table as a giant array. The first URL is at index 0, the next URL is at index 1, the second URL at index 2, and so on. Each user of your application would start by watching the video at URL index 0, then URL index 1, etc. This would ensure the user never sees the same video twice. You would not need to store a list of all the videos they've seen. Instead, you could store the index of the last video they saw.
Your application could grab the first n videos from the table to present to your users. Once that list was exhausted, it could go grab the next n videos. And so on...
What I've described here is essentially how pagination is implemented in DynamoDB. To bring this abstraction back to the world of DynamoDB, your algorithm could look something like this:
Scan the URLArray table for the first "page" of URLs (a scan operation with no filter criteria)
Along with the results, DynamoDB will respond with a LastEvaluatedKey, which will allow you to retrieve the next page of results starting from this position
Present your user with each video you pulled back from the scan operation, making sure to record the id (the Primary Key) the the last video they saw.
When you exhaust the URLs from step 1, execute another scan operation with the ExclusiveStartKey set to the LastEvaluatedKey returned from step 2.
When users return to your application, query for the next page from the URLArray table with ExclusiveStartKey set to the id of the last video they viewed.
This effectively uses the scan operation to search through your URLArray table one page at a time. Your application would effectively be searching the table from top to bottom, keeping track of where each user is at any given time. When a user revisits your application, just start where they left off.
In response to your edit:
If your use case requires the next video to be unpredictable (e.g. no 2 users can predict what video is next), you have a few problems to solve at the same time:
Selecting an item in an unpredictable/random manner
Tracking what a user has already seen
Putting those two requirements together makes for a tricky access pattern. Let's say you have N videos in your table, and the user has viewed N-1 of these videos leaving only one video unseen. If you are fetching your next video randomly and need to ensure it has not yet been seen, how will you find the last unseen video? How many times would you need to guess before you came across the only unseen video? What query/scan operation could you perform that does this in a single request to DDB? I'm not saying it's impossible, it's just...complicated.
I think it's better to generate a strategy that is unpredictable to the user, but predictable to you when it comes to select the next unseen video.
For example, you could pre-calculate a random order of indexes from 1..N ahead of time, which would represent the order you present the videos for a given user. You could go through that list sequentially, keeping track of the last seen index. That way, you'd always know which video was next and that the video hadn't previously been seen by this user. Fetching that video would be a simple query operation to DDB.
You also asked how to find the number of items in DynamoDB. Unfortunately, there is no DynamoDB equivalent of the SQL count operation. The answer to this question is not straightforward. For the benefit of the community (and to get a diverse set of answers), I'd suggest you make a separate question on Stackoverflow regarding the number of items in a DDB table.
I have few tables as shown below
Polls
PollId Question Option
1 What 1
2 Why 4
Updates
UpdateId Text
1 Sleep
2 Play
Polls and updates are just two sample tables (In reality there are more tables like ,photos, videos,links etc). But when a user visit his home (like facebook new feed) he must be displayed with data relevant to him (no such data included in this example). ie I want to select data from all tables with less number of query executions. (ie, I want to present a mixture of datas, ie polls, photos, videos etc )
Currently, I'm fetching only ids and type (ie which table) from all of the tables and gather further data while iterating through this resultset. (ie from c# calling another SqlQuery) .
Is there a way to query the data from whole tables at once? (OUTER JOIN?, UNION?)
Or simply,
How can I select different type of entities at once in a single sql Query?
You could write your query so that you have one long select list for everything you want and it all comes back in one result set but I suspect that wouldn't work too well because you might have varying numbers of different types of items per user.
If you really must have it all in one hit then you can issue multiple queries in one go and get multiple result sets back. To handle this you can use an ADO.Net DataSet. See this SO example (but not the accepted answer - see Vikram Dibyal's answer as that gives a very basic overview of what I think you're asking for).
I won't copy and paste the stuff from the linked thread, just head over and take a look.
I am having 10000 records
in jsp page 32 columns
29 colums i am using and it is slowing the process any other way
Whether slows down because it is checking for every row.
user dont want pagination he needs to show all 10000 not 1000 records
I have a datalist which iterate , within the iterate loop i gave the logic:equal condition
..................................
for changing the bgcolor only i am using the logic:notequal
Any Idea
Does he can read or print them all. What's a wasting of server/client resource and bandwidth!
But in helpless case, you can you javascript to change color to redistribute server processing. Or use good load strategy, for example: not all at once but delay loading (use ajax I assume), load part in view first,..
The are some articles are written in several parts,
for example, I got those articles from IBM developer works:
Distributed data processing with
Hadoop, Part 1:Getting started
Distributed data processing with
Hadoop, Part 2:Going further
Distributed data processing with
Hadoop, Part 3: Application
development
I will index those three articles separately. And some one search certain keywords, it is possible the part3 is on the top of hit whle part1 is on the 32th. Therefor, if I list results page by page, the part1 and part3 will display on different page.
How can I make sure the hitted documents in the same series displayed together?
I guess in SQL, we can use "group by".
I believe what you are asking for is Field Collapsing, which is currently a trunk feature in Solr, and will be incorporated into the next Solr version.
If you want to roll your own, One possible way to do this is:
Add a "series id" field to each document that is a member of a series. You will have to ensure that this gets incremented for every new series.
Make an initial query to Lucene, and get a hit list.
For each hit, check to see if it has a series id; If it does, make another query by the series id in order to retrieve all the members of the series.
An alternative is to store the ids of all the series members in a field inside each member's document.