Presto Verifier¶

Presto Verifier is a tool to run queries and verify correctness. It can be used to test whether a new Presto version produces the correct query results, or to test if pairs of Presto queries have the same semantics.

During each Presto release, Verifier is run to ensure that there is no correctness regression.

Using Verifier¶

In a MySQL database, create the following table and load it with the queries you would like to run:

CREATE TABLE verifier_queries (
    id int(11) unsigned NOT NULL PRIMARY KEY AUTO_INCREMENT,
    suite varchar(256) NOT NULL,
    name varchar(256) DEFAULT NULL,
    control_catalog varchar(256) NOT NULL,
    control_schema varchar(256) NOT NULL,
    control_query text NOT NULL,
    control_username varchar(256) DEFAULT NULL,
    control_password varchar(256) DEFAULT NULL,
    control_session_properties text DEFAULT NULL,
    test_catalog varchar(256) NOT NULL,
    test_schema varchar(256) NOT NULL,
    test_query text NOT NULL,
    test_username varchar(256) DEFAULT NULL,
    test_password varchar(256) DEFAULT NULL,
    test_session_properties text DEFAULT NULL)

Next, create a config.properties file:

source-query.suites=suite
source-query.database=jdbc:mysql://localhost:3306/mydb?user=my_username&password=my_password
control.hosts=127.0.0.1
control.http-port=8080
control.jdbc-port=8080
control.application-name=verifier-test
test.hosts=127.0.0.1
test.http-port=8081
test.jdbc-port=8081
test.application-name=verifier-test
test-id=1

Verifier can run in either query-bank or control-test mode by setting configuration running-mode.

• control-test: this is the default mode. Both the control query and the test query are executed and their result checksums are compared.

• query-bank: in this mode, the control query is skipped and the comparison is done between a saved snapshot result and the test result.

Create a verifier_snapshots table:

CREATE TABLE verifier_snapshots (
    id int(11) unsigned NOT NULL PRIMARY KEY AUTO_INCREMENT,
    suite varchar(256) NOT NULL,
    name varchar(256) NOT NULL DEFAULT '.',
    is_explain BOOLEAN NOT NULL DEFAULT false,
    snapshot json NOT NULL,
    updated_at datetime NOT NULL DEFAULT now(),
    UNIQUE(suite, name, is_explain));

Download presto-verifier-0.290-executable.jar and rename it to verifier.jar. To run the Verifier:

java -Xmx1G -jar verifier.jar verify config.properties

Before running in query-bank mode, snapshots must be saved. Add configurations:

running-mode=query-bank
save-snapshot=true

Run the verifier and the snapshots will be saved to the table verifier_snapshots.

To run in query-bank mode, set save-snapshot=false or just delete it:

running-mode=query-bank
#save-snapshot=true

Verifier Procedures¶

The following steps summarize the workflow of Verifier.

• Importing Source Queries

  • Reads the list of source queries (query pairs with configuration) from the MySQL table.

• Query Pre-processing and Filtering

  • Applies overrides to the catalog, schema, username, and password of each query.

  • Filters queries according to whitelist and blacklist. Whitelist is applied before blacklist.

  • Filters out queries with invalid syntax.

  • Filters out queries not supported for validation. Select, Insert, CreateTableAsSelect, create table and create view are supported.

• Query rewriting

  • Rewrites queries before execution to ensure that production data is not modified.

  • Rewrites Select queries to CreateTableAsSelect

    • Column names are determined by running the Select query with LIMIT 0.

    • Artificial names are used for unnamed columns.

  • Rewrites Insert and CreateTableAsSelect queries to have their table names replaced.

    • Constructs a setup query to create the table necessary for an Insert query.

  • Rewrites function calls according to nondeterministic-function-substitutes if the configuration is set.

• Query Execution

  • Conceptually, Verifier is configured with a control cluster and a test cluster. However, they may be pointed to the same Presto cluster for certain tests. In Query-bank mode, the control cluster is skipped and a saved snapshot is used instead.

  • For each source query, executes the following queries in order.

    • Control setup queries

    • Control query

    • Test setup queries

    • Test query

    • Control and test teardown queries

  • Queries are subject to timeouts and retries.

    • Cluster connection failures and transient Presto failures are retried.

    • Query retries may conceal reliability issues, and therefore Verifier records all occurred Presto query failures, including the retries.

  • Certain query failures are automatically submitted for re-validation, such as partition dropped or table dropped during query.

  • See Failure Resolution for auto-resolving of query failures.

• Results Comparison

  • For Select, Insert, and CreateTableAsSelect queries, results are written into temporary tables.

  • Constructs and runs the checksum queries for both control and test. If running in query-bank mode, Control checksums will be restored from snapshots saved in mysql database.

  • Control checksums will be saved to mysql database if configuration save-snapshot is set to true.

  • Verifies table schema and row count are the same for the control and the test result table.

  • Verifies checksums are matching for each column. See Column Checksums for special handling of different column types.

  • See Determinism for handling of non-deterministic queries.

• Emitting Results

  • Verification results can be exported as JSON, or human readable text.

Column Checksums¶

For each column in the control/test query, one or more columns are generated in the checksum queries.

• Floating Point Columns

  • For DOUBLE and REAL columns, 4 columns are generated for verification:

    • Sum of the finite values of the column

    • NAN count of the column

    • Positive infinity count of the column

    • Negative infinity count of the column

  • Checks if NAN count, positive and negative infinity count matches.

  • Checks the nullity of control sum and test sum.

  • If either control mean or test mean very close 0, checks if both are close to 0.

  • Checks the relative error between control sum and test sum.

• VARCHAR Columns

  • For VARCHAR columns, a simple checksum column is generated for verification using the checksum().

  • If validate-string-as-double is set, seven columns below are generated. If NULL counts are equal before and after casting all values to DOUBLE, apply the floating point validation. Otherwise, check if the simple checksum matches.

    • Checksum

    • Count of the NULL values

    • Count of the NULL values after all values are casted to DOUBLE

    After casting all values to DOUBLE

    • Sum of the finite values

    • NAN count of the values

    • Positive infinity count of the values

    • Negative infinity count of the values

• Array Columns

  • For an array column arr of type array(E), three columns are generated for verification:

    • Sum of the cardinality

    • Checksum of the cardinality

    • Array checksum

      • If E is not orderable, array checksum is checksum(arr).

      • If E is orderable, array checksum is coalesce(checksum(try(array_sort(arr))), checksum(arr)).

  • If use-error-margin-for-floating-point-arrays is set and E is DOUBLE or REAL, these six columns below are generated instead. Check if the sum of the cardinality matches and if the checksum of the cardinality matches. Apply the floating point validation to the rest of results.

    • Sum of the cardinality

    • Checksum of the cardinality

    • Sum of the finite elements of all array values

    • Count of the NAN elements of all array values

    • Count of the positive infinity elements of all array values

    • Count of the negative infinity elements of all array values

  • If validate-string-as-double is set and E is VARCHAR, these nine columns below are generated instead. Check if the sum and the checksum of the cardinality match. If NULL counts are equal before and after casting all array elements to DOUBLE, apply the floating point validation. Otherwise, check if the array checksum matches.

    • Sum of the cardinality

    • Checksum of the cardinality

    • Array checksum checksum(array_sort(arr))

    • Count of the NULL elements of all array values

    • Count of the NULL elements after all array elements are casted to DOUBLE

    After casting all array elements to DOUBLE

    • Sum of the finite elements of all array values

    • Count of the NAN elements of all array values

    • Count of the positive infinity elements of all array values

    • Count of the negative infinity elements of all array values

• Map Columns

  • For a map column of type map(K, V), four columns are generated for verification:

    • Sum of the cardinality

    • Checksum of the map

    • Array checksum of the key set

    • Array checksum of the value set

  • If validate-string-as-double is set and K is VARCHAR, six additional columns are generated:

    • Count of the NULL elements of all key sets

    • Count of the NULL elements of the key sets after all map keys are casted to DOUBLE

    After casting all map keys to DOUBLE

    • Sum of the finite elements of all key sets

    • Count of the NAN elements of all key sets

    • Count of the positive infinity elements of all key sets

    • Count of the negative infinity elements of all key sets

  • If validate-string-as-double is set and V is VARCHAR, six additional columns are generated:

    • Count of the NULL elements of all value sets

    • Count of the NULL elements of the value sets after all map values are casted to DOUBLE

    After casting all map values to DOUBLE

    • Sum of the finite elements of all value sets

    • Count of the NAN elements of all value sets

    • Count of the positive infinity elements of all value sets

    • Count of the negative infinity elements of all value sets

• Row Columns

  • Checksums row fields recursively according to the type of the fields.

For all other column types, generates a simple checksum using the checksum() function.

Determinism¶

A result mismatch, either a row count mismatch or a column mismatch, can be caused by non-deterministic query features. To avoid false alerts, we perform determinism analysis for the control query. If a query is found non-deterministic, we skip the verification as it does not provide insights.

Determinism analysis follows the following steps. If a query is found non-deterministic at any point, the analysis will conclude.

• Non-deterministic catalogs can be specified with determinism.non-deterministic-catalog. If a query references any table from those catalogs, the query is considered non-deterministic.

• Runs the control query again and compares the results with the initial control query run.

• If a query has a LIMIT n clause but no ORDER BY clause at the top level:

  • Runs a query to count the number of rows produced by the control query without the LIMIT clause.

  • If the resulting row count is greater than n, treats the control query as non-deterministic.

Failure Resolution¶

The differences in configuration, including cluster size, can cause a query to succeed on the control cluster but fail on the test cluster. A checksum query can also fail, which may be due to limitation of Presto or Presto Verifier. Thus, we allow Verifier to automatically resolve certain query failures.

• EXCEEDED_GLOBAL_MEMORY_LIMIT: Resolves if the control query uses more memory than the test query.

• EXCEEDED_TIME_LIMIT: Resolves unconditionally.

• TOO_MANY_HIVE_PARTITIONS: Resolves if the test cluster does not have enough workers to make sure the number of partitions assigned to each worker stays within the limit.

• COMPILER_ERROR, GENERATED_BYTECODE_TOO_LARGE: Resolves if the control checksum query fails with this error. If the control query has too many columns, generated checksum queries might be too large in certain cases.

In cases of result mismatches, Verifier may be giving noisy signals, and we allow Verifier to automatically resolve certain mismatches.

• Structured-typed Columns: If array element or map key/value contains floating point types, column checksum is unlikely to match.

  • For an array column, resolve if the element type contains floating point types and the cardinality checksum matches.

  • For a map column, resolve the mismatch when both of the following conditions are true:

    • The cardinality checksum matches.

    • The checksum of the key or value that does not contains floating point types matches.

  • Resolve a test case only when all columns are resolved.

• Resolved Functions: In the case of a results mismatch, if the query uses a function in a

  specified list, the test case is marked as resolved.

Explain Mode¶

In explain mode, Verifier checks whether source queries can be explained instead of whether they produces the same results. Verification is marked as succeeded when both control query and test query can be explained.

The field matchType in the output event can be used as an indicator whether there are plan differences between the control run and the test run.

For non-DML queries, the control query and the plan comparison are skipped.

Extending Verifier¶

Verifier can be extended for further behavioral changes in addition to configuration properties.

AbstractVerifyCommand shows the components that be extended. Implement the abstract class and create a command line wrapper similar to PrestoVerifier.

Configuration Reference¶

General Configuration¶

Query Override Configuration¶

The following configurations control the behavior of query metadata modification before verification starts. Counterparts are also available for test queries with prefix control being replaced with test.

Query Execution Configuration¶

The following configurations control the behavior of query execution on the control cluster. Counterparts are also available for test clusters with prefix control being replaced with test.

Determinism Analyzer Configuration¶

Failure Resolution Configuration¶