I have dataset like this A person identified by an ID, use some object identified by another ID and amount of time he use that. I want to know the first 20 items highly used by the person. Amount of data is very large over 100 million and each id can produce about 200 object he may use.
So first thing i created a projection table with cluster and keep the things sorted how the things will happen in the mapper so that all the things will be at one place in the node so that mapper when it is distributing will find the things locally
CREATE TABLE person_objectid_dwell ( person string, objectid string, sum_dwell bigint)
CLUSTERED BY (person) SORTED BY (sum_dwell desc,objectid asc)INTO 100 BUCKETS STORED AS ORC;
And once done i inserted the data from the feeder table like this
insert into person_objectid_dwell select person, objectid, sum_dwell from person_objectid_dwell distribute by person sort by sum_dwell desc, objectid asc;
And then query using windowing with a table creation
create table person_top20_objectsdwell as select * from ( select person, objectid, sum_dwell,
rank() over (partition by person order by sum_dwell desc ) as rank
from person_objectid_dwell ) t where rank <21;
Problem is this i am not getting the performance what i think i should get, i set the number of reducers etc. Program is running with 3000+ mapper and 1000+ reducers and mapping phase is not getting over at all.
Related
I have a scenario where I have two Hive tables, and the second one is essentially an evolved schema of the first (it has 1 more column in this example).
Table_A
{
business_date String
Name String
Age Number
} partitioned by business_date
Table_B {
business_date String
Name String
Age Number
Address String
} partitioned by business_date
In order to obfuscate downstream users from schema changes, I am creating a Hive view with the following syntax:
Create VIEW customer_info AS
select * from Table_B
UNION
select business_date, name, age, null as address from Table_A
I know the above returns all the data, but from a performance standpoint, if a query run against the view with a valid business_date value, does it take the partition key into account? Or do I lose this benefit when working with views?
Edit: I should mention that business_date is actually a unique value across all partitions. This means, that data provided in Table_A, should not be provided in Table_B. Think of Table_A as being an "older version" of data. Given this, is this the best approach of serving the data if the goal is to abstract schema changes away from the end consumers?
Edit#2: Storing this data in one table is not possible due to tons of other problems.
You are not using any partition predicates in your query, that is why it will be no partition pruning. Use explain command to check this, it will show partition predicates applied. Partition pruning should work fine with a view.
UNION is the same as UNION ALL+DISTINCT.
Use UNION ALL instead if applicable, it will perform much better.
On the other hand, partitioning by something unique will create partitions with single row, this will kill your hive metastore probably. Hope you mean something else saying that
business_date is actually a unique value across all partitions
Remove partitioning in this case and the performance will be significantly better.
I have bucketed table based on column flightnum(10 buckets), data size is approx 700MB, bucketing enforced as well.
When I am executing the query :
select count(flightnum) from flight_buck where flightnum=10;
getting the response in approx 46s. Total number of mappers were 27.
When executing the same query on non bucketed table with same data :
select count(flightnum) from flight_temp where flightnum=10;
getting the response in approx 47s. Total number of mappers used were 30.
Why I am getting the response in same amount of time?
Bucketing helps join to be faster, to increase the simple SELECT speed you have to use partitioned tables.
Try to partition table by flightnum and run again the selects.
Why does this happen ?
Let's create a bucketed not partitioned table like this:
create table `t1b`(
`exchange` string,
`stock_symbol` string,
`date` string,
`stock_price_open` float,
`stock_price_high` float,
`stock_price_low` float,
`stock_price_close` float,
`stock_volume` int,
`stock_price_adj_close` float)
clustered by ( `stock_symbol` ) sorted by ( `date` ) into 306 buckets;
And let's fill it with data... There are as many reducers as many buckets because each reducer will process only record with the same keys and will store the data into its file using the sorting you like, in this case by date
Let's look at HDFS...
Please note what we got.. 306 files (buckets) ...
and inside each of them there are records which have the same clustering key...
But all the files are into the same folder, and when SELECTing with hive there is no way to understand which files hold the value we are looking for, so bucketing with no partitioning does not speed up select because there are no info about where are the data we are looking for.
What does bucketing do ? When you are JOINing data, the whole bucket can be loaded into RAM and we can get a fast join in MAP instead to get a slow join in REDUCE.
I've recently started to play around with Cassandra. My understanding is that in a Cassandra table you define 2 keys, which can be either single column or composites:
The Partitioning Key: determines how to distribute data across nodes
The Clustering Key: determines in which order the records of a same partitioning key (i.e. within a same node) are written. This is also the order in which the records will be read.
Data from a table will always be sorted in the same order, which is the order of the clustering key column(s). So a table must be designed for a specific query.
But what if I need to perform 2 different queries on the data from a table. What is the best way to solve this when using Cassandra ?
Example Scenario
Let's say I have a simple table containing posts that users have written :
CREATE TABLE posts (
username varchar,
creation timestamp,
content varchar,
PRIMARY KEY ((username), creation)
);
This table was "designed" to perform the following query, which works very well for me:
SELECT * FROM posts WHERE username='luke' [ORDER BY creation DESC];
Queries
But what if I need to get all posts regardless of the username, in order of time:
Query (1): SELECT * FROM posts ORDER BY creation;
Or get the posts in alphabetical order of the content:
Query (2): SELECT * FROM posts WHERE username='luke' ORDER BY content;
I know that it's not possible given the table I created, but what are the alternatives and best practices to solve this ?
Solution Ideas
Here are a few ideas spawned from my imagination (just to show that at least I tried):
Querying with the IN clause to select posts from many users. This could help in Query (1). When using the IN clause, you can fetch globally sorted results if you disable paging. But using the IN clause quickly leads to bad performance when the number of usernames grows.
Maintaining full copies of the table for each query, each copy using its own PRIMARY KEY adapted to the query it is trying to serve.
Having a main table with a UUID as partitioning key. Then creating smaller copies of the table for each query, which only contain the (key) columns useful for their own sort order, and the UUID for each row of the main table. The smaller tables would serve only as "sorting indexes" to query a list of UUID as result, which can then be fetched using the main table.
I'm new to NoSQL, I would just want to know what is the correct/durable/efficient way of doing this.
The SELECT * FROM posts ORDER BY creation; will results in a full cluster scan because you do not provide any partition key. And the ORDER BY clause in this query won't work anyway.
Your requirement I need to get all posts regardless of the username, in order of time is very hard to achieve in a distributed system, it supposes to:
fetch all user posts and move them to a single node (coordinator)
order them by date
take top N latest posts
Point 1. require a full table scan. Indeed as long as you don't fetch all records, the ordering can not be achieve. Unless you use Cassandra clustering column to order at insertion time. But in this case, it means that all posts are being stored in the same partition and this partition will grow forever ...
Query SELECT * FROM posts WHERE username='luke' ORDER BY content; is possible using a denormalized table or with the new materialized view feature (http://www.doanduyhai.com/blog/?p=1930)
Question 1:
Depending on your use case I bet you could model this with time buckets, depending on the range of times you're interested in.
You can do this by making the primary key a year,year-month, or year-month-day depending on your use case (or finer time intervals)
The basic idea is that you bucket changes for what suites your use case. For example:
If you often need to search these posts over months in the past, then you may want to use the year as the PK.
If you usually need to search the posts over several days in the past, then you may want to use a year-month as the PK.
If you usually need to search the post for yesterday or a couple of days, then you may want to use a year-month-day as your PK.
I'll give a fleshed out example with yyyy-mm-dd as the PK:
The table will now be:
CREATE TABLE posts_by_creation (
creation_year int,
creation_month int,
creation_day int,
creation timeuuid,
username text, -- using text instead of varchar, they're essentially the same
content text,
PRIMARY KEY ((creation_year,creation_month,creation_day), creation)
)
I changed creation to be a timeuuid to guarantee a unique row for each post creation event. If we used just a timestamp you could theoretically overwrite an existing post creation record in here.
Now we can then insert the Partition Key (PK): creation_year, creation_month, creation_day based on the current creation time:
INSERT INTO posts_by_creation (creation_year, creation_month, creation_day, creation, username, content) VALUES (2016, 4, 2, now() , 'fromanator', 'content update1';
INSERT INTO posts_by_creation (creation_year, creation_month, creation_day, creation, username, content) VALUES (2016, 4, 2, now() , 'fromanator', 'content update2';
now() is a CQL function to generate a timeUUID, you would probably want to generate this in the application instead, and parse out the yyyy-mm-dd for the PK and then insert the timeUUID in the clustered column.
For a usage case using this table, let's say you wanted to see all of the changes today, your CQL would look like:
SELECT * FROM posts_by_creation WHERE creation_year = 2016 AND creation_month = 4 AND creation_day = 2;
Or if you wanted to find all of the changes today after 5pm central:
SELECT * FROM posts_by_creation WHERE creation_year = 2016 AND creation_month = 4 AND creation_day = 2 AND creation >= minTimeuuid('2016-04-02 5:00-0600') ;
minTimeuuid() is another cql function, it will create the smallest possible timeUUID for the given time, this will guarantee that you get all of the changes from that time.
Depending on the time spans you may need to query a few different partition keys, but it shouldn't be that hard to implement. Also you would want to change your creation column to a timeuuid for your other table.
Question 2:
You'll have to create another table or use materialized views to support this new query pattern, just like you thought.
Lastly if your not on Cassandra 3.x+ or don't want to use materialized views you can use Atomic batches to ensure data consistency across your several de-normalized tables (that's what it was designed for). So in your case it would be a BATCH statement with 3 inserts of the same data to 3 different tables that support your query patterns.
The solution is to create another tables to support your queries.
For SELECT * FROM posts ORDER BY creation;, you may need some special column for grouping it, maybe by month and year, e.g. PRIMARY KEY((year, month), timestamp) this way the cassandra will have a better performance on read because it doesn't need to scan the whole cluster to get all data, it will also save the data transfer between nodes too.
Same as SELECT * FROM posts WHERE username='luke' ORDER BY content;, you must create another table for this query too. All column may be same as your first table but with the different Primary Key, because you cannot order by the column that is not the clustering column.
I have a table that stores millions of url, date and name entries. Each row is unique in terms of either:
url + date
or
date + name.
I require this table to be stored in descending date order so that when I query it I can simply "SELECT * FROM mytable LIMIT 1000" to get me the most recent 1000 records, no sorting involved. Does anyone know how to set things up to do this please? To the best of my current understanding I am trying the following but it does not store them in date order:
CREATE TABLE mytable (
url text,
date timestamp,
name text,
PRIMARY KEY ((url, name), date)
)
WITH CLUSTERING ORDER BY (date DESC);
To store the data according to an order, you'd need to change the partitioner to byte ordered. This is no longer a good idea...it's maintained for back compat, but there are issues:
http://www.datastax.com/documentation/cassandra/2.1/cassandra/architecture/architecturePartitionerBOP_c.html
You could also apply bucketing and query over your buckets. Each bucket being a partition, and each partition would have data stored in order. Not exactly what you want, but worth trying.
I recently was reading about Oracle Index Organized Tables (IOTs) but am not sure I quite understand WHEN to use them. So I have a small table:
create table categories
(
id VARCHAR2(36),
group VARCHAR2(100),
category VARCHAR2(100
)
create unique index (group, category, id) COMPRESS 2;
The id column is a foreign key from another table entries and my common query is:
select e.id, e.time, e.title from entries e, categories c where e.id=c.id AND e.group=? AND c.category=? ORDER by e.time
The entries table is indexed properly.
Both of these tables have millions (16M currently) of rows and currently this query really stinks (note: I have it wrapped in a pagination query also so I only get back the first 20, but for simplicity I omitted that).
Since I am basically indexing the entire table, does it make sense to create this table as an IOT?
EDIT by popular demand:
create table entries
(
id VARCHAR2(36),
time TIMESTAMP,
group VARCHAR2(100),
title VARCHAR2(500),
....
)
create index (group, time) compress 1;
My real question I dont think depends on this though. Basically if you have a table with few columns (3 in this example) and you are planning on putting a composite index on all three rows is there any reason not to use an IOT?
IOTs are great for a number of purposes, including this case where you're gonna have an index on all (or most) of the columns anyway - but the benefit only materialises if you don't have the extra index - the idea is that the table itself is an index, so put the columns in the order that you want the index to be in. In your case, you're accessing category by id, so it makes sense for that to be the first column. So effectively you've got an index on (id, group, category). I don't know why you'd want an additional index on (group, category, id).
Your query:
SELECT e.id, e.time, e.title
FROM entries e, categories c
WHERE e.id=c.id AND e.group=? AND c.category=?
ORDER by e.time
You're joining the tables by ID, but you have no index on entries.id - so the query is probably doing a hash or sort merge join. I wouldn't mind seeing a plan for what your system is doing now to confirm.
If you're doing a pagination query (i.e. only interested in a small number of rows) you want to get the first rows back as quick as possible; for this to happen you'll probably want a nested loop on entries, e.g.:
NESTED LOOPS
ACCESS TABLE BY ROWID - ENTRIES
INDEX RANGE SCAN - (index on ENTRIES.group,time)
ACCESS TABLE BY ROWID - CATEGORIES
INDEX RANGE SCAN - (index on CATEGORIES.ID)
Since the join to CATEGORIES is on ID, you'll want an index on ID; if you make it an IOT, and make ID the leading column, that might be sufficient.
The performance of the plan I've shown above will be dependent on how many rows match the given "group" - i.e. how selective an average "group" is.
Have you looked at dba-oracle.com, asktom.com, IOUG, another asktom.com?
There are penalties to pay for IOTs - e.g., poorer insert performance
Can you prototype it and compare performance?
Also, perhaps you might want to consider a hash cluster.
IOT's are a trade off. You are getting access performance for decreased insert/update performance. We typically use them for reference data that is batch loaded daily and not updated during the day. This is not to say it's the only way to use them, just how we use them.
Few things here:
You mention pagination - have you considered the first_rows hint?
Is that the order your index is in, with group as the first field? If so I'd consider moving ID to be the first column since that index will not be used.
foreign keys should have an index on the column. Consider addind an index on the foreign key (id column).
Are you sure it's not the ORDER BY causing slowness?
What version of Oracle are you using?
I ASSUME there is a primary key on table entries for field id, correct?
Why the WHERE condition does not include "c.group = e.group" ?
Try to:
Remove the order by condition
Change the index definition from "create unique index (group,
category, id)" to "create unique index (id, group, category)"
Reorganise table categories as an IOT on (group, category, id)
Reorganise table categories as an IOT on (id, group, category)
In each of the above case use EXPLAIN PLAN to review the cost