The MySQL connector uses several threads, so previously upon connector shutdown these threads were simply cancelled. This is fine for the binlog reader (which can stop at any moment), but is a poor approach for the snapshot as we didn’t always properly release the database resources and also didn’t complete the writing of the DDL history.
With this change, the snapshot reader stops in a very controlled manner, basically by having the 10-step snapshot procedure frequently check whether the reader is to continue working, and to completely avoid thread interruption altogether. And, the snapshot procedure will always clean up its database resources (locks, transactions, etc.), even if the procedure is stopped before completion.
This change also refactors how the snapshot and binlog reader are managed. This is no longer done in the MySqlConnectorTask class (which is busy enough), but rather the logic has been encapsulated in a new `ChainedReader` that makes use of a new `Reader` interface. This makes testing of `ChainedReader` easier, and ensure that `ChainedReader` relies only upon the primary methods of `Reader` rather than upon `AbstractReader`. `ChainedReader` handles multiple readers generically, and ensures that when stopped the readers are all handled correctly and completely process all records, yet avoid accidentally starting a subsequent reader(s) when stopping the previous reader.
The MySQL DDL parser was not properly consuming function declarations. For functions, the parser consumes the entire statement without handline the various expressions within the function declaration, but the parser was not properly finding the end of the statement and instead was continuing to try to consume values beyond the end of the statement.
Specifically, when the parser consumes a `BEGIN`, it looks for a corresponding `END`. However, if it encountered an `END IF`, the `IF` plus any remaining tokens were left on the token stream and unprocessed. This confused the parser, which keep looking for statements and ultimately ended with a `No more content` error.
This case was replicated in integration tests, and the code fixed to properly find the end of the statements.
By default the MySQL connector handles `DECIMAL` and `NUMERIC` columns using `java.math.BigDecimal` values and describing them using the `org.apache.kafka.connect.data.Decimal` schema type, which serializes the values to a binary form.
This change adds a configuration option that will keep the default behavior, but will instead allow handling `DECIMAL` adn `NUMERIC` values as Java `double` and a schema type of `FLOAT64`.
Added tests to verify whether the connector is properly restarting in the binlog when previously the connector failed or stopped in the middle of a transaction. The tests showed that the connector is not able to properly start when using or not using GTIDs, since restarting from an arbitrary binlog event causes problems since the TABLE_MAP events for the affected tables are skipped.
The logic was changed significantly to record in the offsets the binlog coordinates at the start of the transaction, which should work whether or not GTIDs are used. Upon restart, the connector may have to re-read the events that were previously processed, but now the offset also includes the number of events that were previously processed so that these can be skipped upon restart.
This has an unforunate side effect since the offsets capture a transaction was completed only when it generates a source record for the subsequent transaction. This is because the connector generates source records (with their offsets) for the binlog events in the transaction before the transaction's commit is seen. And, since no additional source records are produced for the transaction commit, the recorded offsets will show that the prior transaction is complete and that all of the events in the subsequent transaction are to be skipped. Thus, upon restart the connector has to re-read (but ignore) all of the binlog events associated with the completed transaction. This shouldn’t be a problem, and will only slow restarts for very large transactions.
Improved the error handling of the MySQL connector to ensure that we’re always stopping the connector when we have a problem handling a binlog event or if we have problems starting.
When the MySQL connector is reading the binlog, it outputs INFO log messages reporting status at an exponentially-increasing rate, starting at every 5 seconds and doubling until a max period of 1 hour. This output is useful when the connector starts to know that it is working, but thereafter the usefulness decreases. Once an hour is probably acceptable output.
This is not intended to replace the capturing of metrics, but is merely an aid to easily tell via the logs whether the connector continues to work.
Also improved the log message when the binlog reader stops to capture the total number of events recorded by Kafka Connect and the last recorded offset.
Corrected how the MySQL connector is treating columns of type `BIT(n)`, where _n_ is the number of bits in the value. When `n=1`, the resulting values are booleans; when `n>1`, the resulting values are little endian `byte[]` that have the minimum number of bytes to hold the `n` bits.
The `KafkaDatabaseHistory` was always creating a new producer whenever its `start()` method was called, even if it were called more than once. And, the `MySqlSchema` was calling `start()` twice, resulting in multiple producers being created and registered with JMX. Both issues were fixed.
Also, UUIDs were being used as the name of the JMX MBean for the producer, unless the `database.history.consumer.client.id` and `database.history.producer.client.id` properties were being explicitly set. Now, the MySQL connector will by default set the `client.id` property on both the database history's Kafka consumer and producer to `{connectorName}-dbhistory`. Of course, the `database.history.consumer.client.id` and `database.history.producer.client.id` properties can still be set to define the name of the producer and consumer.
MySQL supports "zero-value" dates and timestamps, but these cannot be represented as valid dates or timestamps using the Java types. For example, the zero-value `0000-00-00` for a date has what Java considers to be an invalid month and day-of-the-month.
This commit changes how the MySQL connector handles these values to not throw exceptions. When columns allow nulls, such values will be treated as nulls; when columns do not allow null values, these values will be converted to a "zero-value" for the corresponding Java representation (e.g., the epoch day or timestamp). A new test case verifies the behaviors.
Anytime we `toString()` a `Configuration`, any values for password properties should be masked. A password property is defined to be a property whose key ends in "password" in a case-insensitive manner.
The MySQL binlog events contain the binary representation of string-like values as encoded per the column's character set. Properly decoding these into Java strings requires capturing the column, table, and database character set when parsing the DDL statements.
Unfortunately, MySQL DDL allows columns (at the time the columns are created or modified) to inherit the default character set for the table, or if that is not defined the default character set for the database, or if that is not defined the character set for the server. So, in addition to modifying the MySQL DDL parser to support capturing the character set name for each column, it also had to be changed to know what these default character set names are.
The default character sets are all available via MySQL server/session/local variables. Although strictly speaking the character set variables cannot be set globally, MySQL DDL does allow session and local variables to be set with `SET` statements. Therefore, this commit enhances the MySQL DDL parser to parse `SET` statements and to track the various global, session, and local variables as seen by the DDL parser. Upon connector startup, a subset of server variables (related to character sets and collations) are read from the database via JDBC and used to initialize the DDL parser via `SET` methods.
In addition to initializing the DDL parser with the system variables related to character sets and collation, it is important to also capture the server and database default character sets in the database history so that the correct character sets are used for columns even when the default character sets have changed on the database and/or the server. Therefore, upon startup or snapshot the MySQL connector records in the database history a `SET` statement for the `character_set_server` and `collation_server` system variables so that, upon a later restart, the history's DDL statements can be re-parsed with the correct default server and database character sets. Also, when the MySQL connector reloads the database history (upon startup), the recorded default server character set is compared with the MySQL instance's current server character set, and if they are different the current character set is recorded with a new `SET` statement.
These extra steps ensure that the connector use the correct character set for each column, even when the connector restarts and reloads the database history captured by a previous version of the connector. IOW, the MySQL connector can be safely upgraded, and the new version will correctly start using the columns' character sets to decode the string-like values.
The DDL parser and in-memory models of the relational schemas were changed to capture the character set for each column whose type is a string (e.g., `CHAR`, `VARCHAR`, etc.). This required handling `SET` statements used to change the system variables that hold the names of the default character set for the server and for each database. So, even if a column does not explicitly define the character set, the column's actual character set is identified from the table's character set, which might default to the current database's character set, which if not set defaults to the system character set.
These changes merely affect how MySQL DDL is parsed and the in-memory relational schema representation to accommodate the character set at various levels. It does not change the behavior of the MySQL connector; that will be done in a subsequent commit.
All tests pass with these changes, including quite a few additional tests for the new functionality.
Changed the MySQL connector to have several new configuration properties for setting up the SSL key store and trust store (which can be used in place of System or JDK properties) used for MySQL secure connections, and another property to specify what kind of SSL connection be used.
Modified several integration tests to ensure all MySQL connections are made with `useSSL=false`.
The ENUM and SET values read from the binlog contain the indexes of the options that are included in the value, but this doesn't compared with the string values returned by MySQL and JDBC that contain the comma-separated options. With this change, the values read from the binlog will also be comma-separated strings.
Rewrote how the MySQL connector converts temporal values to use schemas with names that identify the semantic
type of temporal value, and customized how the MySQL binlog client library creates Java object values from the
raw binlog events.
Several new "semantic" schema types were defined:
* `io.debezium.time.Year` represents a year number as an INT32 value (e.g., 2016, -345, etc.).
* `io.debezium.time.Date` represents a date by storing the epoch seconds (that is, the number of seconds past the epoch) as an INT64 value.
* `io.debezium.time.Time` represents a time by storing the milliseconds past midnight as an INT32 value.
* `io.debezium.time.MicroTime` represents a time by storing the microsconds past midnight as an INT32 value.
* `io.debezium.time.NanoTime` represents a time by storing the nanoseconds past midnight as an INT32 value.
* `io.debezium.time.Timestamp` represents a date and time (without timezone information) by storing the milliseconds past epoch as an INT64 value.
* `io.debezium.time.MicroTimestamp` represents a date and time (without timezone information) by storing the microseconds past epoch as an INT64 value.
* `io.debezium.time.NanoTimestamp` represents a date and time (without timezone information) by storing the nanoseconds past epoch as an INT64 value.
* `io.debezium.time.ZonedTime` represents a time with timezone and optional fractions of a second (but no date) by storing the ISO8601 form as a STRING value (e.g., `10:15:30+01:00`)
* `io.debezium.time.ZonedTimestamp` represents a date and time with timezone and optional fractions of a second by storing the ISO8601 form as a STRING value (e.g., `2011-12-03T10:15:30.030431+01:00`)
This range of semantic types allows for a far more accurate representation in the events of the temporal values stored within the database. The MySQL connector chooses the semantic type based upon the precision of the MySQL type (e.g., `TIMESTAMP(6)` will be represented with `io.debezium.time.MicroTimestamp`, whereas `TIMESTAMP(3)` will be represented with `io.debezium.time.Timestamp`). This ensures that the events do not lose precision and that the semantics of the database column values are retained in the events even though the values are represented with primitive values.
Obviously these Kafka Connect schema representations are different and more precise than the built-in `org.apache.kafka.connect.data.Date`, `org.apache.kafka.connect.data.Time`, and `org.apache.kafka.connect.data.Timestamp` logical types provided by Kafka Connect and used by the MySQL connector in all 0.2.x and 0.1.x versions. Migration to the new MySQL connector should be possible, although consumers may still need to know about these types to properly handle temporal values and the correct precision (i.e., consumers can just assume all date INT64 values represent milliseconds).
The MySQL binlog client library converted the raw binary event information to JDBC types using a local Calendar instance, which obviously incorporates the local timezone and cannot retain more than millisecond precision. This change extends the library's deserializers to instead use the Java 8 `javax.time` classes and to retain the exact semantics of the database values and to not lose any precisions (since the `javax.time` classes have nanosecond precision).
The same logic is also used to convert the JDBC values obtained during a snapshot from the MySQL Connect/J JDBC driver. The latter has a few quirks, such as not returning any fractional seconds for `TIME` columns, even though `java.sql.Time` can store up to milliseconds.
Most of the logic of the conversions of values and mapping to Kafka Connect schemas is handled in the new `JdbcValueConverters`, which was extracted from the existing `TableSchemaBuilder`. The MySQL connector reuses and actually extends the `JdbcValueConverters` class with its own `MySqlValueConverters` class that also adds support for MySQL-specific types such as `YEAR`. Other connectors whose values are based on JDBC types should be able to reuse and/or extend the `JdbcValueConverters` class.
Integration tests that deal with temporal types were modified to use proper expected values and comparisons.
By default the MySQL JDBC driver will put the entire result set into memory, which obviously doesn't work for tables of even moderate sizes. This change adds support for streaming rows in result sets when the tables have more than a configurable number of rows (defaults to 1,000).
This posed a problem for how we were previously finding the last row in the last table; the MySQL driver does not support `ResultSet.isLast()` on result sets that are streamed. Instead, this commit wraps the consumer to which the snapshot reader writes all source records, with a consumer that buffers the last record. When the snapshot completes, the offset is updated (denoting the end of the snapshot) and set on the last buffered record before that record is flushed to the normal consumer. This should add minimal overhead while simplifying the logic to ensure the last source record has the updated offset.
This also improves the log output of the snapshot process.
* fixes a java.sql.Date conversion test to take into account zone offsets
* makes sure the ZK DB is closed during testing, otherwise file handles may leak and cause test failures
Upgraded from Kafka 0.9.0.1 to Kafka 0.10.0. The only required change was to override the `Connector.config()` method, which returns `null` or a `ConfigDef` instance that contains detailed metadata for each of the configuration fields, including supporting recommended values and marking fields as not visible (e.g., if they don't make sense given other configuration field values). This can be used by user interfaces to data-drive the configuration of a connector. Also, the default validation logic of the Connector implementations uses a `Validator` that is pretty restrictive in its functionality.
Debezium already had a fairly decent and simple `Configuration` framework. After several attempts to try and merge these concepts, reconciling the two validation mechanisms was very complicated and involved a lot of changes. It was easier to simply continue Debezium-specific validation and to override the `Connector.validate(...)` method to use Debezium's `Configuration`-based validation. Connector-based validation logic includes determining recommended values, so Debezium's `Field` class (used to define each configuration property) was enhanced with a new `Recommender` class that is similar to Kafka's.
Additional integration tests were added to verify that the `ConfigDef` result is acceptable and that the new connector validation logic works as expected, including getting recommended values for some fields (e.g., database names, table/collection names) from MySQL and MongoDB by connecting and dynamically reading the values. This was done in a way that remains backward compatible with the regular expression formats of these fields, but in a user interface that uses the `ConfigDef` mechanism the user can simply select the databases and table/collection identifiers.
Added an integration test case to diagnose the loss of the fractional seconds from MySQL temporal values. The problem appears to be a bug in the MySQL Binary Log Connector library that we used, and this bug was reported as https://github.com/shyiko/mysql-binlog-connector-java/issues/103. That was fixed in version 0.3.2 of the library, which Stanley was kind enough to release for us.
During testing, though, several issues were discovered in how temporal values are handled and converted from the MySQL events, through the MySQL Binary Log client library, and through the Debezium MySQL connector to conform with Kafka Connect's various temporal logical schema types. Most of the issues involved converting most of the temporal values from local time zone (which is how they are created by the MySQL Binary Log client) into UTC (which is how Kafka Connect expects them). Really, java.util.Date doesn't have time zone information and instead tracks the number of milliseconds past epoch, but the conversion of normal timestamp information to the milliseconds past epoch in UTC depends on the time zone in which that conversion happens.
The MySQL connector now maps TINYINT and SMALLINT columns to INT16 (rather than INT32) because INT16 is smaller and yet still large enough for all TINYINT and SMALLINT values. Note that the range of TINYINT values is either -128 to 127 for signed or 0 to 255 for unsigned, and thus INT8 is not an acceptable choice since it can only handle values in the range 0 to 255. Additionally, the JDBC Specification also suggests the proper Java type for SQL-99's TINYINT is short, which maps to Kafka Connect's INT16.
This change will be backward compatible, although the generated Kafka Connect schema will be different than in previous versions. This shouldn't cause a problem, since clients should expect to handle schema changes, and this schema change does comply with Avro schema evolution rules.
Added a new `debezium-connector-mongodb` module that defines a MongoDB connector. The MongoDB connector can capture and record the changes within a MongoDB replica set, or when seeded with addresses of the configuration server of a MongoDB sharded cluster, the connector captures the changes from the each replica set used as a shard. In the latter case, the connector even discovers the addition of or removal of shards.
The connector monitors each replica set using multiple tasks and, if needed, separate threads within each task. When a replica set is being monitored for the first time, the connector will perform an "initial sync" of that replica set's databases and collections. Once the initial sync has completed, the connector will then begin tailing the oplog of the replica set, starting at the exact point in time at which it started the initial sync. This equivalent to how MongoDB replication works.
The connector always uses the replica set's primary node to tail the oplog. If the replica set undergoes an election and different node becomes primary, the connector will immediately stop tailing the oplog, connect to the new primary, and start tailing the oplog using the new primary node. Likewise, if connector experiences any problems communicating with the replica set members, it will try to reconnect (using exponential backoff so as to not overwhelm the replica set) and continue tailing the oplog from where it last left off. In this way the connector is able to dynamically adjust to changes in replica set membership and to automatically handle communication failures.
The MongoDB oplog contains limited information, and in particular the events describing updates and deletes do not actually have the before or after state of the documents. Instead, the oplog events are all idempotent, so updates contain the effective changes that were made during an update, and deletes merely contain the deleted document identifier. Consequently, the connector is limited in the information it includes in its output events. Create and read events do contain the initial state, but the update contain only the changes (rather than the before and/or after states of the document) and delete events do not have the before state of the deleted document. All connector events, however, do contain the local system timestamp at which the event was processed and _source_ information detailing the origins of the event, including the replica set name, the MongoDB transaction timestamp of the event, and the transactions identifier among other things.
It is possible for MongoDB to lose commits in specific failure situations. For exmaple, if the primary applies a change and records it in its oplog before it then crashes unexpectedly, the secondary nodes may not have had a chance to read those changes from the primary's oplog before the primary crashed. If one such secondary is then elected as primary, it's oplog is missing the last changes that the old primary had recorded and no longer has those changes. In these cases where MongoDB loses changes recorded in a primary's oplog, it is possible that the MongoDB connector may or may not capture these lost changes.
