Does creating a Spark DataFrame and saving it in Parquet format guarantee that the order of columns in the parquet file will be preserved?
Ex) A Spark DataFrame is created with columns A, B, C, and saved as Parquet. When the Parquet files are read, will the column order always be A, B, C?
I've noticed that if I save a Spark DataFrame, and then read the parquet files, the column order is preserved:
df.select("A", "B", "C").write.save(...)
df = spark.read.load(...)
df.printSchema()
root
|-- A: string (nullable = true)
|-- B: string (nullable = true)
|-- C: string (nullable = true)
Then, if I save by selecting a different order of columns, and then read the parquet files, I can see the order is also what I expect:
df.select("C", "B", "A").write.save(...)
df = spark.read.load(...)
df.printSchema()
root
|-- C: string (nullable = true)
|-- B: string (nullable = true)
|-- A: string (nullable = true)
However, I can't seem to find any documentation supporting this, and the comments of this post: Is there a possibility to keep column order when reading parquet? have conflicting information.
Related
I have two Parquet directories that are being loaded into Spark. They have all the same columns, but the columns are in a different order.
val df = spark.read.parquet("url1").printSchema()
root
|-- a: string (nullable = true)
|-- b: string (nullable = true)
val df = spark.read.parquet("url2").printSchema()
root
|-- b: string (nullable = true)
|-- a: string (nullable = true)
val urls = Array("url1", "url2")
val df = spark.read.parquet(urls: _*).printSchema()
root
|-- a: string (nullable = true)
|-- b: string (nullable = true)
When I load the files together they always seem to take on the ordering of url1. I'm worried that having the parquet files in url1 and url2 saved in a different order will have unintended consequences, such as a and b swapping values. Can someone explain how parquet loads columns stored in a different order, with links to official documentation, if possible?
I am trying to read orc file of a managed hive table using below pyspark code.
spark.read.format('orc').load('hive managed table path')
when i do a print schema on fetched dataframe, it is as follow
root
|-- operation: integer (nullable = true)
|-- originalTransaction: long (nullable = true)
|-- bucket: integer (nullable = true)
|-- rowId: long (nullable = true)
|-- currentTransaction: long (nullable = true)
|-- row: struct (nullable = true)
| |-- col1: float (nullable = true)
| |-- col2: integer (nullable = true)
|-- partition_by_column: date (nullable = true)
Now i am not able to parse this data and do any manipulation on data frame. While applying action like show(), i am getting an error saying
java.lang.IllegalArgumentException: Include vector the wrong length
did someone face the same issue? if yes can you please suggest how to resolve it.
It's a known issue.
You get that error because you're trying to read Hive ACID table but Spark still doesn't have support for this.
Maybe you can export your Hive table to normal ORC files and then read them with Spark or try using alternatives like Hive JDBC as described here
As i am not sure about the versions You can try other ways to load the ORC file.
Using SqlContext
val df = sqlContext.read.format("orc").load(orcfile)
OR
val df= spark.read.option("inferSchema", true).orc("filepath")
OR SparkSql(recommended)
import spark.sql
sql("SELECT * FROM table_name").show()
I'm trying to convert json files to parquet with very few transformations (adding date) but I then need to partition this data before saving it to parquet.
I'm hitting a wall on this area.
Here is the creation process of the table:
df_temp = spark.read.json(data_location) \
.filter(
cond3
)
df_temp = df_temp.withColumn("date", fn.to_date(fn.lit(today.strftime("%Y-%m-%d"))))
df_temp.createOrReplaceTempView("{}_tmp".format("duration_small"))
spark.sql("CREATE TABLE IF NOT EXISTS {1} LIKE {0}_tmp LOCATION '{2}/{1}'".format("duration_small","duration", warehouse_location))
spark.sql("DESC {}".format("duration"))
then regarding the save of the conversion:
df_final.write.mode("append").format("parquet").partitionBy("customer_id", "date").saveAsTable('duration')
but this generates the following error:
pyspark.sql.utils.AnalysisException: '\nSpecified partitioning does not match that of the existing table default.duration.\nSpecified partition columns: [customer_id, date]\nExisting partition columns: []\n ;'
the schema being:
root
|-- action_id: string (nullable = true)
|-- customer_id: string (nullable = true)
|-- duration: long (nullable = true)
|-- initial_value: string (nullable = true)
|-- item_class: string (nullable = true)
|-- set_value: string (nullable = true)
|-- start_time: string (nullable = true)
|-- stop_time: string (nullable = true)
|-- undo_event: string (nullable = true)
|-- year: integer (nullable = true)
|-- month: integer (nullable = true)
|-- day: integer (nullable = true)
|-- date: date (nullable = true)
Thus I tried to change the create table to:
spark.sql("CREATE TABLE IF NOT EXISTS {1} LIKE {0}_tmp PARTITIONED BY (customer_id, date) LOCATION '{2}/{1}'".format("duration_small","duration", warehouse_location))
But this create an error like:
...mismatched input 'PARTITIONED' expecting ...
So I discovered that PARTITIONED BY doesn't work with LIKE but I'm running out of ideas.
if using USING instead of LIKE I got the error:
pyspark.sql.utils.AnalysisException: 'It is not allowed to specify partition columns when the table schema is not defined. When the table schema is not provided, schema and partition columns will be inferred.;'
How am I supposed to add a partition when creating the table?
Ps - Once the schema of the table is defined with the partitions, I want to simply use:
df_final.write.format("parquet").insertInto('duration')
I finally figured out how to do it with spark.
df_temp.read.json...
df_temp.createOrReplaceTempView("{}_tmp".format("duration_small"))
spark.sql("""
CREATE TABLE IF NOT EXISTS {1}
USING PARQUET
PARTITIONED BY (customer_id, date)
LOCATION '{2}/{1}' AS SELECT * FROM {0}_tmp
""".format("duration_small","duration", warehouse_location))
spark.sql("DESC {}".format("duration"))
df_temp.write.mode("append").partitionBy("customer_id", "date").saveAsTable('duration')
I don't know why, but if I can't use insertInto, it uses a weird customer_id out of nowhere and doesn't append the different dates.
I have a orc file, when read with below option it reads all the columns .
val df= spark.read.orc("/some/path/")
df.printSChema
root
|-- id: string (nullable = true)
|-- name: string (nullable = true)
|-- value: string (nullable = true)
|-- all: string (nullable = true)
|-- next: string (nullable = true)
|-- action: string (nullable = true)
but I want to read only two columns from that file , is there any way to read only two columns (id,name) while loading orc file ?
is there any way to read only two columns (id,name) while loading orc file ?
Yes, all you need is subsequent select. Spark will take care of the rest for you:
val df = spark.read.orc("/some/path/").select("id", "name")
Spark has lazy execution model. So you can do any data transformation in you code without immediate real effect. Only after action call Spark start to doing job. And Spark are smart enough not to do extra work.
So you can write like this:
val inDF: DataFrame = spark.read.orc("/some/path/")
import spark.implicits._
val filteredDF: DataFrame = inDF.select($"id", $"name")
// any additional transformations
// real work starts after this action
val result: Array[Row] = filteredDF.collect()
I'm performing an inner join between dataframes to only keep the sales for specific days:
val days_df = ss.createDataFrame(days_array.map(Tuple1(_))).toDF("DAY_ID")
val filtered_sales = sales.join(days_df,Seq("DAY_ID")
filtered_sales.show()
This results in an empty filtered_sales dataframe (0 records), both columns DAY_ID have the same type (string).
root
|-- DAY_ID: string (nullable = true)
root
|-- SKU: string (nullable = true)
|-- DAY_ID: string (nullable = true)
|-- STORE_ID: string (nullable = true)
|-- SALES_UNIT: integer (nullable = true)
|-- SALES_REVENUE: decimal(20,5) (nullable = true)
The sales df is populated from a 20GB file.
Using the same code with a small file of some KB will work fine with the join and I can see the results. The empty result dataframe occurs only with bigger dataset.
If I change the code and use the following one, it works fine even with the 20GB sales file:
sales.filter(sales("DAY_ID").isin(days_array:_*))
.show()
What is wrong with the inner join?
Try to broadcast days_array and then apply inner join. As days_array is too small compared to another table, broadcasting will help.