I have a number of hive warehouses. The data resides in parquet files in Amazon S3. Some of the tables contain TB of data. Currently in hive most tables are partitioned by a combination of month and year, both of which are saved mainly as string. Other fields are either bigint, int, float, double, string and unix timestamps. Our goal is to migrate them to apache iceberg tables. The challenge is how to choose the partition keys.
I have already calculated the cardinality of each field in each table by:
Select COUNT(DISTINCT my_column) As my_column_count
From my_table;
I have also calculated the percentage of null values for each field:
SELECT 100.0 * count(*)/number_of_all_records
FROM my_db.my_table
Where my_column IS NULL;
In short I already know three things for each field:
Data type
Cardinality
Percentage of null values
By knowing these three pieces of information, my question is how should I choose the best column or combination of columns as partition keys for my future iceberg tables? Are there any rule of thumbs?
How many partitions is considered as optimized when choosing partition keys? What data type is best when choosing partition keys? What are other factors that need to be considered? Is bucketing the same in iceberg tables as it is in hive and how it can be leveraged by the partition keys? What data types are best for partition keys? Is it better to have many small partitions or having a few big partitions? Any other aspects in partition keys that need to be considered?
One crucial part is missing from your description - the queries. You need to understand what are the queries that will run on this data. Understanding the queries that will run on the data (to the best you can) is super important.
For example, consider a simple table with: Date, Id, Name, Age as columns.
If the queries are date based meaning, it will query the data in the context of dates,
select * from table where date > 'some-date'
then it's a good idea to partition by date.
However, if the queries are age related
select * from table where age between 20 and 30
then you should consider partition by age or age groups
I have come across Hive tables which I need to convert to Redshift/MySql equivalent.
I am having trouble understanding Hive query structure and would appreciate some help:
CREATE TABLE IF NOT EXISTS table_1 (
id BIGINT,
price DOUBLE,
asset string
)
PARTITIONED BY (
pt STRING
);
ALTER TABLE table_1 DROP IF EXISTS PARTITION (pt== '${yyyymmdd}');
INSERT OVERWRITE TABLE table_1 PARTITION (pt= '${yyyymmdd}')
select aa.id,aa.price,aa.symbol from
...
...
from
table_2 table
I am having trouble understanding the PARTITIONED BY clause. This, if I am understanding correctly, is different from MySQL table partitions, and is a Hive specific dynamic partition.
The partition does not define a column or a key, and partitions by the current date.
Does this mean that table_1 is partitioned by the date? Each day has a separate partition?
Then later on in the code there are notations similar to
inner join table_new table on table.pt = '${yyyymmdd}' and ...
In this context, does it mean only rows inserted on yyyymmdd are selected for the join?
Thank you.
Partition in Hive is a folder in HDFS by default with name key=value + metadata in the Hive metastore. You can alter partition location and create partition on top of any folder.
This PARTITIONED BY (pt STRING) defines partition column pt of type string, not date. Partition values are stored in the metadata. The pt column is not present in the table data files, it is only defined in PARTITIONED BY, all partition values are stored in the metadata. If you load partition dynamically, partition folder is being created with name pt='value'.
This sentence creates partition dynamically:
INSERT OVERWRITE TABLE table_1 PARTITION (pt)
select id, price, symbol
coln as pt --partition column should be the last one
from ...
And this sentence loads single STATIC partition:
INSERT OVERWRITE TABLE table_1 PARTITION (pt= '${yyyymmdd}')
select aa.id,aa.price,aa.symbol
from
No partition column is selected, partition value specified in the
PARTITION (pt= '${yyyymmdd}')
'${yyyymmdd}' here is a parameter with name yyyymmdd which is passed to the script using --hivevar like this:
hive --hivevar yyyymmdd=20200604 -f myscript.sql
You can pass ANY string as partition value in this case, though parameter name yyyymmdd suggests it's format.
BTW date format in hive is 'yyyy-MM-dd' Strings in 'yyyy-MM-dd' format can be implicitly converted to DATE.
I will try in one shot explain what is partitioning in Hive. First of all would be
WHEN TO USE TABLE PARTITIONING
Table partitioninig is good when:
Reading the entire dataset takes too long
Queries almost always filter on the partition columns
There are a reasonable number of different values for partition columns
Data generation of ETL process splits data by file or directory names
Partition column values are not in the data itself
Don't partition on columns with many unique values
Example: Partitioning customers by first name
CREATING PARTITIONED TABLES
To create a partitioned table, use the PARTITIONED BY clause in the CREATE TABLE statement.
The names and types of the partition columns must be specified
in the PARTITIONED BY clause, and only in the PARTITIONED BY clause.
They must not also appear in the list of all the other columns.
CREATE TABLE customers_by_country
(cust_id STRING, name STRING)
PARTITIONED BY (country STRING)
ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t';
The example CREATE TABLE statement shown above creates the table customers_by_country,
which is partitioned by the STRING column named country.
Notice that the country column appears only in the PARTITIONED BY clause,
and not in the column list above it.
This example specifies only one partition column, but you can specify more than one by using
a comma-separated column list in the PARTITIONED BY clause.
Aside from these specific differences, this CREATE TABLE statement is the same
as the statement used to create an equivalent non-partitioned table.
Table partitioning is implemented in a way that is mostly transparent
to a user issuing queries with Hive.
A partition column is what’s known as a virtual column, because its values are not stored within the data files.
Following is the result of the DESCRIBE command on customers_by_country;
it displays the partition column country just as if it were a normal column within the table.
You can refer to partition columns in any of the usual clauses of a SELECT statement.
name type comment
cust_id string
name string
country string
You can load data in partitioned tables dynamically or statically
LOADING DATA WITH DYNAMIC PARTITION
One way to load data into a partitioned table is to use dynamic partitioning,
which automatically defines partitions when you load the data, using the values in the partition column.
(The other way is to manually define the partitions with Static Partitioning)
To use dynamic partitioning, you must load data using an INSERT statement.
In the INSERT statement, you must use the PARTITION clause to list the partition columns.
The data you are inserting must include values for the partition columns.
The partition columns must be the rightmost columns in the data you are inserting,
and they must be in the same order as they appear in the PARTITION clause.
INSERT OVERWRITE TABLE customers_by_country
PARTITION(country)
SELECT cust_id, name, country FROM customers;
The example shown above uses an INSERT … SELECT statement
to load data into the customers_by_country table with dynamic partitioning.
Notice that the partition column, country, is included
in the PARTITION clause and is specified last in the SELECT list.
When Hive executes this statement, it automatically creates partitions
for the country column and loads the data into these partitions based on the values in the country column.
The resulting data files in the partition subdirectories do not include values for the country column.
Since the country is known based on which subdirectory a data file is in,
it would be redundant to include country values in the data files as well.
Look at the contents of the customers_by_country directory.
It should now have one subdirectory for each value in the country column.
Look at the file in one of those directories.
Notice that the file contains the row for the customer from that country,
and no others; notice also that the country value is not included.
Note: Hive includes a safety feature that prevents users
from accidentally creating or overwriting a large number of partitions.
(See “Risks of Using Partitioning” for more about this.)
By default, Hive sets the property hive.exec.dynamic.partition.mode to strict.
This prevents you from using dynamic partitioning, though you can still use static partitions.
You can disable this safety feature in Hive by setting
the property hive.exec.dynamic.partition.mode to nonstrict:
SET hive.exec.dynamic.partition.mode=nonstrict;
Then you can use the INSERT statement to load the data dynamically.
Hive properties set in Beeline are for the current session only,
so the next time you start a Hive session this property will be set back to strict.
But you or your system administrator can configure properties permanently, if necessary.
When you run some SELECT queries on the partitioned table, if the table is big enough you can note significant difference in the time it takes to run.
Notice that you will not query the table any differently than you would query the customers table.
LOADING DATA WITH STATIC PARTITIONING
One way to load data into a partitioned table is to use static partitioning,
in which you manually define the different partitions.
With static partitioning, you create a partition manually, using an ALTER TABLE … ADD PARTITION statement,
and then load the data into the partition.
For example, this ALTER TABLE statement creates the partition for Pakistan (pk):
ALTER TABLE customers_by_country
ADD PARTITION (country='pk');
Notice how the partition column name, which is country, and the specific value that defines this partition,
which is pk, are both specified in the ADD PARTITION clause.
This creates a partition directory named country=pk inside the customers_by_country table directory.
After the partition for Pakistan is created, you can add data into the partition using an INSERT … SELECT statement:
INSERT OVERWRITE TABLE customers_by_country
PARTITION(country='pk')
SELECT cust_id, name FROM customers WHERE country='pk'
Notice how in the PARTITION clause, the partition column name, which is country,
and the specific value, which is pk, are both specified, just like in the ADD PARTITION command used to create the partition.
Also notice that in the SELECT statement, the partition column is not included in the SELECT list.
Finally, notice that the WHERE clause in the SELECT statement selects only customers from Pakistan.
With static partitioning, you need to repeat these two steps for each partition:
first create the partition, then add data.
You can actually use any method to load the data; you need not use an INSERT statement.
You could instead use hdfs dfs commands or a LOAD DATA INPATH command.
But however you load the data, it’s your responsibility to ensure that data is stored in the correct partition subdirectories.
For example, data for customers in Pakistan must be stored in the Pakistan partition subdirectory,
and data for customers in other countries must be stored in those countries’ partition subdirectories.
Static partitioning is most useful when the data being loaded
into the table is already divided into files based on the partition column,
or when the data grows in a manner that coincides with the partition column:
For example, suppose your company opens a new store in a different country,
like New Zealand ('nz'), and you're given a file of data for new customers, all from that country.
You could easily add a new partition and load that file into it.
RISKS OF USING PARTITIONING
A major risk when using partitioning is creating partitions that lead you into the small files problem.
When this happens, partitioning a table will actually worsen query performance
(the opposite of the goal when using partitioning) because it causes too many small files to be created.
This is more likely when using dynamic partitioning, but it could still
happen with static partitioning—for example if you added a new partition to a sales table
on a daily basis containing the sales from the previous day,
and each day’s data is not particularly big.
When choosing your partitions, you want to strike a happy balance between too many partitions
(causing the small files problem) and too few partitions (providing performance little benefit).
The partition column or columns should have a reasonable number of values
for the partitions—but what you should consider reasonable is difficult to quantify.
Using dynamic partitioning is particularly dangerous because if you're not careful,
it's easy to partition on a column with too many distinct values.
Imagine a use case where you are often looking for data that falls within
a time frame that you would specify in your query.
You might think that it's a good idea to partition on a column that pertains to time.
But a TIMESTAMP column could have the time to the nanosecond, so every row could have a unique value;
that would be a terrible choice for a partition column! Even to the minute or hour could create
far too many partitions, depending on the nature of your data;
partitioning by larger time units like day, month, or even year might be a better choice.
As another example, consider an employees table.
This has five columns: empl_id, first_name, last_name, salary, and office_id.
Before reading on, think for a moment, which of these might be reasonable for partitioning
The column empl_id is a unique identifier.
If that were your partition column, you would have a separate partition for each employee,
and each would have exactly one row.
In addition, it's not likely you'll be doing a lot of queries looking for a particular value,
or even a particular range of values. This is a poor choice.
The column first_name will not have one per employee, but there will likely be many columns that have only one row.
This is also true for last_name.
Also, like empl_id, it's not likely you'll need filter queries based on these columns. These are also poor choices.
The column salary also will have many divisions
(and even more so if your salaries go to the cent rather than to the dollar as our sample table does).
While it may be that you'll sometimes want to query on salary ranges,
it's not likely you'll want to use individual salaries.
So salary is a poor choice.
A more limited salary_grades specification, like the ones in the salary_grades table,
might be reasonable if your use case involves looking at the data by salary grade frequently.
The office_id column identifies the office where an employee works.
This will have a much smaller number of unique values, even if you have a large company with offices in many cities.
It's imaginable that your use case might be to frequently filter
your employee data based on office location, too. So this would be a good choice.
You also can use multiple columns and create nested partitions.
For example, a dataset of customers might include country and state_or_province columns.
You can partition by country and then partition those further by state_or_province, so customers from Ontario,
Canada would be in the country=ca/state_or_province=on/ partition directory.
This can be extremely helpful for large amounts of data that you want to access either by country or by state or province.
However, using multiple columns increases the danger of creating too many partitions, so you must take extra care when doing so.
The risk of creating too many partitions is why Hive includes the property
hive.exec.dynamic.partition.mode, set to strict by default, which must be reset to nonstrict before you can create a partition.
Rather than automatically and mechanically resetting that property when you're about to load data dynamically,
take it as an opportunity to think about the partitioning columns
and maybe check the number of unique values you would get when you load the data.
And that's all.
When I am trying to load data from existing Partitioned table to new Clustered table in Bigquery, I see below error.
Failed to copy Natural partitioned table to Natural partitioned clustering table: not supported
Any inputs please.
As per https://cloud.google.com/bigquery/docs/creating-clustered-tables
you can create the table as :
CREATE TABLE
mydataset.myclusteredtable
PARTITION BY
DATE(timestamp)
CLUSTER BY
clustering_column AS
SELECT
*
FROM
`mydataset.mytable`
NOTE: Only standard SQL is supported for querying clustered tables and for writing query results to clustered tables.
Note, BigQuery supports few kinds of partitions:
Ingestion time: Tables are partitioned based on the data's ingestion (load) date or arrival date.
Date/timestamp: Tables are partitioned based on a TIMESTAMP or DATE column.
Integer range: This feature is in a pre-release state
Tables can only be partitioned by one field, which must be a timestamp or date column, and clustered by a single set of columns.
Prerequisites:
When performing this process, you must have the permissions outlined in BigQuery’s documentation.
When you use the clustered table feature with a partitioned table, you are subject to the limits on partitioned tables.
Quotas and limits also apply to the different types of jobs you can run against clustered tables, including loading data (load jobs), what happens in your case.
Take a look for creating clustered tables in BigQuery.
If all of the source tables are partitioned tables, the partition specifications for all source tables must match the destination table's partition specification. The source and destination tables must be in datasets in the same location.
I have a partitioned table with name like mytable_* where the suffix denote date.
I would now like to convert this to the new way that tables are to be partitioned in bigquery, ie. with the _partitiondate column, etc.
I was thinking of creating the new table's schema based on the old one, and then to insert data into it, but I am not sure how to put the date value (that is the suffix) of the old table into the _partitiondate field.
If you have previously created date-sharded tables, you can convert the entire set of related tables into a single ingestion-time partitioned table by using the partition command in the bq command-line tool. The date-sharded tables must use the following naming convention: [TABLE]_YYYYMMDD. For example, mytable_20160101, ... , mytable_20160331.
For this you should use bq partition command like in example below
bq --location=[LOCATION] partition --time_partitioning_type=DAY --time_partitioning_expiration [INTEGER] [PROJECT_ID]:[DATASET].[SOURCE_TABLE]_ [PROJECT_ID]:[DATASET].[DESTINATION_TABLE]
You can see more details / options in below articles
Converting date-sharded tables into ingestion-time partitioned tables
and
bq partition
im planning to build a new ads system and we are considering to use google bigquery.
ill quickly describe my data flow :
Each User will be able to create multiple ADS. (1 user, N ads)
i would like to store the ADS impressions and i thought of 2 options.
1- create a table for impressions , for example table name is :Impressions fields : (userid,adsid,datetime,meta data fields...)
in this options of all my impressions will be stored in a single table.
main pros : ill be able to big data queries quite easily.
main cons: table will be hugh, and with multiple queries, ill end up paying too much (:
option 2 is to create table per ads
for example, ads id 1 will create
Impression_1 with fields (datetime,meta data fields)
pros: query are cheaper, data table is smaller
cons: todo big dataquery sometimes ill have to create a union and things will complex
i wonder what are your thoughts regarding this ?
In BigQuery it's easy to do this, because you can create tables per each day, and you have the possibility to query only those tables.
And you have Table wildcard functions, which are a cost-effective way to query data from a specific set of tables. When you use a table wildcard function, BigQuery only accesses and charges you for tables that match the wildcard. Table wildcard functions are specified in the query's FROM clause.
Assuming you have some tables like:
mydata.people20140325
mydata.people20140326
mydata.people20140327
You can query like:
SELECT
name
FROM
(TABLE_DATE_RANGE(mydata.people,
TIMESTAMP('2014-03-25'),
TIMESTAMP('2014-03-27')))
WHERE
age >= 35
Also there are Table Decorators:
Table decorators support relative and absolute <time> values. Relative values are indicated by a negative number, and absolute values are indicated by a positive number.
To get a snapshot of the table at one hour ago:
SELECT COUNT(*) FROM [data-sensing-lab:gartner.seattle#-3600000]
There is also TABLE_QUERY, which you can use for more complex queries.