Table Deletes, Updates, and Merges

Delta Lake supports several statements to facilitate deleting data from and updating data in Delta tables.

Delete from a table

You can remove data that matches a predicate from a Delta table. For instance, to delete all events from before 2017, you can run the following:

Scala

import io.delta.tables._

val deltaTable = DeltaTable.forPath(spark, "/data/events/")

deltaTable.delete("date < '2017-01-01'")        // predicate using SQL formatted string

import org.apache.spark.sql.functions._
import spark.implicits._

deltaTable.delete(col("date") < "2017-01-01")       // predicate using Spark SQL functions and implicits

Java

import io.delta.tables.*;
import org.apache.spark.sql.functions;

DeltaTable deltaTable = DeltaTable.forPath(spark, "/data/events/");

deltaTable.delete("date < '2017-01-01'");            // predicate using SQL formatted string

deltaTable.delete(functions.col("date").lt(functions.lit("2017-01-01")));   // predicate using Spark SQL functions

Python

from delta.tables import *
from pyspark.sql.functions import *

deltaTable = DeltaTable.forPath(spark, "/data/events/")

deltaTable.delete("date < '2017-01-01'")        # predicate using SQL formatted string

deltaTable.delete(col("date") < "2017-01-01")   # predicate using Spark SQL functions

See Delta Lake API Reference for more details.

Important

delete removes the data from the latest version of the Delta table but does not remove it from the physical storage until the old versions are explicitly vacuumed. See vacuum for more details.

Tip

When possible, provide predicates on the partition columns for a partitioned Delta table as such predicates can significantly speed up the operation.

Update a table

You can update data that matches a predicate in a Delta table. For example, to fix a spelling mistake in the eventType, you can run the following:

Scala

import io.delta.tables._

val deltaTable = DeltaTable.forPath(spark, "/data/events/")

deltaTable.updateExpr(            // predicate and update expressions using SQL formatted string
  "eventType = 'clck'",
  Map("eventType" -> "'click'")

import org.apache.spark.sql.functions._
import spark.implicits._

deltaTable.update(                // predicate using Spark SQL functions and implicits
  col("eventType") === "clck",
  Map("eventType" -> lit("click")));

Java

import io.delta.tables.*;
import org.apache.spark.sql.functions;
import java.util.HashMap;

DeltaTable deltaTable = DeltaTable.forPath(spark, "/data/events/");

deltaTable.updateExpr(            // predicate and update expressions using SQL formatted string
  "eventType = 'clck'",
  new HashMap<String, String>() {{
    put("eventType", "'click'");
  }}
);

deltaTable.update(                // predicate using Spark SQL functions
  functions.col(eventType).eq("clck"),
  new HashMap<String, Column>() {{
    put("eventType", functions.lit("click"));
  }}
);

Python

from delta.tables import *
from pyspark.sql.functions import *

deltaTable = DeltaTable.forPath(spark, "/data/events/")

deltaTable.update("eventType = 'clck'", { "eventType": "'click'" } )   # predicate using SQL formatted string

deltaTable.update(col("eventType") == "clck", { "eventType": lit("click") } )   # predicate using Spark SQL functions

See Delta Lake API Reference for more details.

Tip

Similar to delete, update operations can get a significant speedup with predicates on partitions.

Upsert into a table using Merge

You can upsert data from an Apache Spark DataFrame into a Delta table using the merge operation. This operation is similar to the SQL MERGE command but has additional support for deletes and extra conditions in updates, inserts, and deletes.

Suppose you have a Spark DataFrame that contains new data for events with eventId. Some of these events may already be present in the events table. To merge the new data into the events table, you want to update the matching rows (that is, eventId already present) and insert the new rows (that is, eventId not present). You can run the following:

Scala

import io.delta.tables._
import org.apache.spark.sql.functions._

val updatesDF = ...  // define the updates DataFrame[date, eventId, data]

DeltaTable.forPath(spark, "/data/events/")
  .as("events")
  .merge(
    updatesDF.as("updates"),
    "events.eventId = updates.eventId")
  .whenMatched
  .updateExpr(
    Map("data" -> "updates.data"))
  .whenNotMatched
  .insertExpr(
    Map(
      "date" -> "updates.date",
      "eventId" -> "updates.eventId",
      "data" -> "updates.data"))
  .execute()

Java

import io.delta.tables.*;
import org.apache.spark.sql.functions;
import java.util.HashMap;

Dataset<Row> updatesDF = ...  // define the updates DataFrame[date, eventId, data]

DeltaTable.forPath(spark, "/data/events/")
  .as("events")
  .merge(
    updatesDF.as("updates"),
    "events.eventId = updates.eventId")
  .whenMatched()
  .updateExpr(
    new HashMap<String, String>() {{
      put("data" -> "events.data");
    }})
  .whenNotMatched()
  .insertExpr(
    new HashMap<String, String>() {{
      put("date", "updates.date");
      put("eventId", "updates.eventId");
      put("data", "updates.data");
    }})
  .execute();

Python

from delta.tables import *

deltaTable = DeltaTable.forPath(spark, "/data/events/")

deltaTable.alias("events").merge(
    updatesDF.alias("updates"),
    "events.eventId = updates.eventId") \
  .whenMatchedUpdate(set = { "data" : "updates.data" } ) \
  .whenNotMatchedInsert(values =
    {
      "date": "updates.date",
      "eventId": "updates.eventId",
      "data": "updates.data"
    }
  ) \
  .execute()

Here is a detailed description of the merge programmatic operation.

  • There can be 1, 2, or 3 when[Matched | NotMatched] clauses. Of these, at most 2 can be whenMatched clauses, and at most 1 can be a whenNotMatched clause.
  • whenMatched clause:
    • Can have at most one update action and one delete action.
    • Can have an optional condition. However, if there are two clauses, then the first one must have a condition.
    • When there are two clauses and there are conditions (or the lack of) such that a row matches both clauses, the first clause/action is executed. In other words, the order of the clauses matters.
    • If none of the clauses matches a source-target row pair that satisfy the merge condition, the target rows are not updated.
    • To update all the columns of the target Delta table with the corresponding column of the source DataFrame, use whenMatched(...).updateAll(). This is equivalent to updateExpr(Map("col1" -> "source.col1", "col2" -> "source.col2", ...)).
  • whenNotMatched clause:
    • Can have only the insert action, which can have an optional condition.
    • If not present or if it is present but the non-matching source row does not satisfy the condition, the source row is not inserted.
    • To insert all the columns of the target Delta table with the corresponding column of the source Dataframe, use whenMatched(...).insertAll(). This is equivalent to insertExpr(Map("col1" -> "source.col1", "col2" -> "source.col2", ...)).
  • See the Delta Lake API Reference for details.

Important

A merge operation can fail if multiple rows of the source dataset match and attempt to update the same same row of the target Delta table. By SQL semantics of merge, such an update operation is ambiguous as it is unclear which source should be used to update the matching target row. You can preprocess the source table to eliminate the possibility of multiple matches. See the change-data-capture example – it preprocesses the change dataset (that is, the source dataset) to retain only the latest change for each key before applying that change into the target Delta table.

Tip

You should add as much information to the merge condition to both reduce the amount of work and the chance of transaction conflicts. For example, suppose you have a table that is partitioned by country and date and you want to use merge to update information for the last day country by country. Adding the condition events.date = current_date() AND events.country = 'USA' will make the query faster as it will look for matches only in the relevant partitions.

Merge examples

Here are a few examples on how to use merge in different scenarios.

Data deduplication when writing into Delta tables

A common ETL use case is to collect logs into Delta table by appending them to a table. However, often the sources can generate duplicate log records and downstream deduplication steps are needed to take care of them. With merge, you can avoid inserting the duplicate records.

Scala

deltaTable
  .as("logs")
  .merge(
    newDedupedLogs.as("newDedupedLogs"),
    "logs.uniqueId = newDedupedLogs.uniqueId")
  .whenNotMatched()
  .insertAll()
  .execute()

Python

deltaTable.alias("logs").merge(
    newDedupedLogs.alias("newDedupedLogs"),
    "logs.uniqueId = newDedupedLogs.uniqueId") \
  .whenNotMatchedInsertAll() \
  .execute()

Note

  • The dataset containing the new logs needs to be deduplicated within itself. By the SQL semantics of merge, it will match and deduplicated the new data with the existing data in the table, but if there is duplicate data within the new dataset, it will get inserted. Hence, deduplicate the new data before merging into the table.

  • If you know that you may get duplicate records only for a few days, you can optimized your query further by partitioning the table by date, and then specifying the date range of the target table to match on.

    Scala

    deltaTable.as("logs").merge(
        newDedupedLogs.as("newDedupedLogs"),
        "logs.uniqueId = newDedupedLogs.uniqueId AND logs.date > current_date() - INTERVAL 7 DAYS")
      .whenNotMatched("newDedupedLogs.date > current_date() - INTERVAL 7 DAYS")
      .insertAll()
      .execute()
    

    Python

    deltaTable.alias("logs").merge(
        newDedupedLogs.alias("newDedupedLogs"),
        "logs.uniqueId = newDedupedLogs.uniqueId AND logs.date > current_date() - INTERVAL 7 DAYS") \
      .whenNotMatchedInsertAll("newDedupedLogs.date > current_date() - INTERVAL 7 DAYS") \
      .execute()
    

    This will be more efficient than the previous command as it will looks for duplicates only in the last 7 days of logs, not the entire table.

  • You can use this insert-only merge with Structured Streaming to perform continuous deduplication of the logs.
    • In a streaming query, you can use merge operation in foreachBatch to continuously write any streaming data to a Delta table with deduplication. See the streaming example below for more information on foreachBatch.
    • In another streaming query, you can continuously read deduplicated data from this Delta table. This is possible because insert-only merge will only append new data to the Delta table.

Slowly changing data (SCD) Type 2 operation into Delta tables

Another common operation is SCD Type 2, which maintains history of all changes made to each key in a dimensional table. Such operations require updating existing rows to mark previous values of keys as old, and the inserting the new rows as the latest values. Given a source table with updates and the target table with the dimensional data, SCD Type 2 can be expressed with merge.

Here is a concrete example of maintaining the history of addresses for a customer along with the active date range of each address. When a customer’s address needs to be updated, you have to mark the previous address as not the current one, update its active date range, and add the new address as the current one.

Scala

val customersTable: DeltaTable = ...   // table with schema (customerId, address, current, effectiveDate, endDate)

val updatesDF: DataFrame = ...          // DataFrame with schema (customerId, address, effectiveDate)

// Rows to INSERT new addresses of existing customers
val newAddressesToInsert = updatesDF
  .as("updates")
  .join(customersTable.toDF.as("customers"), "customerid")
  .where("customers.current = true AND updates.address <> customers.address")

// Stage the update by unioning two sets of rows
// 1. Rows that will be inserted in the whenNotMatched clause
// 2. Rows that will either update the current addresses of existing customers or insert the new addresses of new customers
val stagedUpdates = newAddressesToInsert
  .selectExpr("NULL as mergeKey", "updates.*")   // Rows for 1.
  .union(
    updatesDF.selectExpr("updates.customerId as mergeKey", "*")  // Rows for 2.
  )

// Apply SCD Type 2 operation using merge
customersTable
  .as("customers")
  .merge(
    stagedUpdates.as("staged_updates"),
    "customers.customerId = mergeKey")
  .whenMatched("customers.current = true AND customers.address <> staged_updates.address")
  .updateExpr(Map(                                      // Set current to false and endDate to source's effective date.
    "current" -> "false",
    "endDate" -> "staged_updates.effectiveDate"))
  .whenNotMatched()
  .insertExpr(Map(
    "customerid" -> "staged_updates.customerId",
    "address" -> "staged_updates.address",
    "current" -> "true",
    "effectiveDate" -> "staged_updates.effectiveDate",  // Set current to true along with the new address and its effective date.
    "endDate" -> "null"))
  .execute()

Python

customersTable = ...  # DeltaTable with schema (customerId, address, current, effectiveDate, endDate)

updatesDF = ...       # DataFrame with schema (customerId, address, effectiveDate)

# Rows to INSERT new addresses of existing customers
newAddressesToInsert = updatesDF \
  .alias("updates") \
  .join(customersTable.toDF().alias("customers"), "customerid") \
  .where("customers.current = true AND updates.address <> customers.address")

# Stage the update by unioning two sets of rows
# 1. Rows that will be inserted in the whenNotMatched clause
# 2. Rows that will either update the current addresses of existing customers or insert the new addresses of new customers
stagedUpdates = (
  newAddressesToInsert
  .selectExpr("NULL as mergeKey", "updates.*")   # Rows for 1
  .union(updatesDF.selectExpr("updates.customerId as mergeKey", "*"))  # Rows for 2.
)

# Apply SCD Type 2 operation using merge
customersTable.alias("customers").merge(
  stagedUpdates.alias("staged_updates"),
  "customers.customerId = mergeKey") \
.whenMatchedUpdate(
  condition = "customers.current = true AND customers.address <> staged_updates.address",
  set = {                                      # Set current to false and endDate to source's effective date.
    "current": "false",
    "endDate": "staged_updates.effectiveDate"
  }
).whenNotMatchedInsert(
  values = {
    "customerid": "staged_updates.customerId",
    "address": "staged_updates.address",
    "current": "true",
    "effectiveDate": "staged_updates.effectiveDate",  # Set current to true along with the new address and its effective date.
    "endDate": "null"
  }
).execute()

Write change data into a Delta table

Similar to SCD, another common use case, often called change data capture (CDC), is to apply all data changes generated from an external database into a Delta table. In other words, a set of updates, deletes, and inserts applied to an external table needs to be applied to a Delta table. You can do this using merge as follows.

Scala

val deltaTable: DeltaTable = ... // DeltaTable with schema (key, value)

// DataFrame with changes having following columns
// - key: key of the change
// - time: time of change for ordering between changes (can replaced by other ordering id)
// - newValue: updated or inserted value if key was not deleted
// - deleted: true if the key was deleted, false if the key was inserted or updated
val changesDF: DataFrame = ...

// Find the latest change for each key based on the timestamp
// Note: For nested structs, max on struct is computed as
// max on first struct field, if equal fall back to second fields, and so on.
val latestChangeForEachKey = changesDF
  .selectExpr("key", "struct(time, newValue, deleted) as otherCols" )
  .groupBy("key")
  .agg(max("otherCols").as("latest"))
  .selectExpr("key", "latest.*")

deltaTable.as("t")
  .merge(
    latestChangeForEachKey.as("s"),
    "s.key = t.key")
  .whenMatched("s.deleted = true")
  .delete()
  .whenMatched()
  .updateExpr(Map("key" -> "s.key", "value" -> "s.newValue"))
  .whenNotMatched("s.deleted = false")
  .insertExpr(Map("key" -> "s.key", "value" -> "s.newValue"))
  .execute()

Python

deltaTable = ... # DeltaTable with schema (key, value)

# DataFrame with changes having following columns
# - key: key of the change
# - time: time of change for ordering between changes (can replaced by other ordering id)
# - newValue: updated or inserted value if key was not deleted
# - deleted: true if the key was deleted, false if the key was inserted or updated
changesDF = spark.table("changes")

# Find the latest change for each key based on the timestamp
# Note: For nested structs, max on struct is computed as
# max on first struct field, if equal fall back to second fields, and so on.
latestChangeForEachKey = changesDF \
  .selectExpr("key", "struct(time, newValue, deleted) as otherCols") \
  .groupBy("key") \
  .agg(max("otherCols").alias("latest")) \
  .select("key", "latest.*") \

deltaTable.alias("t").merge(
    latestChangeForEachKey.alias("s"),
    "s.key = t.key") \
  .whenMatchedDelete(condition = "s.deleted = true") \
  .whenMatchedUpdate(set = {
    "key": "s.key",
    "value": "s.newValue"
  }) \
  .whenNotMatchedInsert(
    condition = "s.deleted = false",
    values = {
      "key": "s.key",
      "value": "s.newValue"
    }
  ).execute()

Upsert from streaming queries using foreachBatch

You can use a combination of merge and foreachBatch (see foreachbatch for more information) to write complex upserts from a streaming query into a Delta table. For example:

  • Write streaming aggregates in Update Mode: This is much more efficient than Complete Mode.

    Scala

    import io.delta.tables.*
    
    val deltaTable = DeltaTable.forPath(spark, "/data/aggregates")
    
    // Function to upsert microBatchOutputDF into Delta table using merge
    def upsertToDelta(microBatchOutputDF: DataFrame, batchId: Long) {
      deltaTable.as("t")
        .merge(
          microBatchOutputDF.as("s"),
          "s.key = t.key")
        .whenMatched().updateAll()
        .whenNotMatched().insertAll()
        .execute()
    }
    
    // Write the output of a streaming aggregation query into Delta table
    streamingAggregatesDF.writeStream
      .format("delta")
      .foreachBatch(upsertToDelta _)
      .outputMode("update")
      .start()
    

    Python

    from delta.tables import *
    
    deltaTable = DeltaTable.forPath(spark, "/data/aggregates")
    
    # Function to upsert microBatchOutputDF into Delta table using merge
    def upsertToDelta(microBatchOutputDF, batchId):
      deltaTable.alias("t").merge(
          microBatchOutputDF.alias("s"),
          "s.key = t.key") \
        .whenMatchedUpdateAll() \
        .whenNotMatchedInsertAll() \
        .execute()
    }
    
    # Write the output of a streaming aggregation query into Delta table
    streamingAggregatesDF.writeStream \
      .format("delta") \
      .foreachBatch(upsertToDelta) \
      .outputMode("update") \
      .start()
    
  • Write a stream of database changes into a Delta table: The merge query for writing change data can be used in foreachBatch to continuously apply a stream of changes to a Delta table.

  • Write a stream data into Delta table with deduplication: The insert-only merge query for deduplication can be used in foreachBatch to continuously write data (with duplicates) to a Delta table with automatic deduplication.

Note

  • Make sure that your merge statement inside foreachBatch is idempotent as restarts of the streaming query can apply the operation on the same batch of data multiple times.
  • When merge is used in foreachBatch, the input data rate of the streaming query (reported through StreamingQueryProgress and visible in the notebook rate graph) may be reported as a multiple of the actual rate at which data is generated at the source. This is because merge reads the input data multiple times causing the input metrics to be multiplied. If this is a bottleneck, you can cache the batch DataFrame before merge and then uncache it after merge.