The main goal of this feature is to enable a command that will
1) Generate a parameterized command for each edit that is in the session
2) Execute that command against the server
3) Update the cached results of the table/view that's being edited with the committed changes (including computed/identity columns)
There's some secret sauce in here where I cheated around worrying about gaps in the updated results. This was accomplished by implementing an IComparable for row edit objects that ensures deletes are the *last* actions to occur and that they occur from the bottom of the list up (highest row ID to lowest). Thus, all other actions that are dependent on the row ID are performed first, then the largest row ID is deleted, then next largest, etc. Nevertheless, by the end of a commit the associated ResultSet is still the source of truth. It is expected that the results grid will need updating once changes are committed.
Also worth noting, although this pull request supports a "many edits, one commit" approach, it will work just fine for a "one edit, one commit" approach.
* WIP
* Adding basic commit support. Deletions work!
* Nailing down the commit logic, insert commits work!
* Updates work!
* Fixing bug in DbColumnWrapper IsReadOnly setting
* Comments
* ResultSet unit tests, fixing issue with seeking in mock writers
* Unit tests for RowCreate commands
* Unit tests for RowDelete
* RowUpdate unit tests
* Session and edit base tests
* Fixing broken unit tests
* Moving constants to constants file
* Addressing code review feedback
* Fixes from merge issues, string consts
* Removing ad-hoc code
* fixing as per @abist requests
* Fixing a couple more issues
This is a very large change. I'll try to outline what's going on.
1. This adds the **EditDataService** which manages editing **Sessions**.
1. Each session has a **ResultSet** (from the QueryExecutionService) which has the rows of the table and basic metadata about the columns
2. Each session also has an **IEditTableMetadata** implementation which is derived from SMO metadata which provides more in-depth and trustworthy data about the table than SqlClient alone can.
3. Each session holds a list of **RowEditBase** abstract class implementations
1. **RowUpdate** - Update cells in a row (generates `UPDATE` statement)
2. **RowDelete** - Delete an entire row (generates `DELETE` statement)
3. **RowCreate** - Add a new row (generates `INSERT INTO` statement)
4. Row edits have a collection of **CellUpdates** that hold updates for individual cells (except for RowDelete)
1. Cell updates are generated from text
5. RowEditBase offers some baseline functionality
1. Generation of `WHERE` clauses (which can be parameterized)
2. Validation of whether a column can be updated
2. New API Actions
1. edit/initialize - Queries for the contents of a table/view, builds SMO metadata, sets up a session
2. edit/createRow - Adds a new RowCreate to the Session
3. edit/deleteRow - Adds a new RowDelete to the Session
4. edit/updateCell - Adds a CellUpdate to a RowCreate or RowUpdate in the Session
5. edit/revertRow - Removes a RowCreate, RowDelete, or RowUpdate from the Session
6. edit/script - Generates a script for the changes in the Session and stores to disk
7. edit/dispose - Removes a Session and releases the query
3. Smaller updates (unit test mock improvements, tweaks to query execution service)
**There are more updates planned -- this is just to get eyeballs on the main body of code**
* Initial stubs for edit data service
* Stubbing out update management code
* Adding rudimentary dispose request
* More stubbing out of update row code
* Adding complete edit command contracts, stubbing out request handlers
* Adding basic implementation of get script
* More in progress work to implement base of row edits
* More in progress work to implement base of row edits
* Adding string => object conversion logic and various cleanup
* Adding a formatter for using values in scripts
* Splitting IMessageSender into IEventSender and IRequestSender
* Adding inter-service method for executing queries
* Adding inter-service method for disposing of a query
* Changing edit contract to include the object to edit
* Fully fleshing out edit session initialization
* Generation of delete scripts is working
* Adding scripter for update statements
* Adding scripting functionality for INSERT statements
* Insert, Update, and Delete all working with SMO metadata
* Polishing for SqlScriptFormatter
* Unit tests and reworked byte[] conversion
* Replacing the awful and inflexible Dictionary<string, string>[][] with a much better test data set class
* Fixing syntax error in generated UPDATE statements
* Adding unit tests for RowCreate
* Adding tests for the row edit base class
* Adding row delete tests
* Adding RowUpdate tests, validation for number of key columns
* Adding tests for the unit class
* Adding get script tests for the session
* Service integration tests, except initialization tests
* Service integration tests, except initialization tests
* Adding messages to sr.strings
* Adding messages to sr.strings
* Fixing broken unit tests
* Adding factory pattern for SMO metadata provider
* Copyright and other comments
* Addressing first round of comments
* Refactoring EditDataService to have a single method for handling
session-dependent operations
* Refactoring Edit Data contracts to inherit from a Session and Row
operation params base class
* Copyright additions
* Small tweak to strings
* Updated unit tests to test the refactors
* More revisions as per pull request comments
Adding a new setting to query execution setting that will change the display value we generate for `BIT` columns. The new setting is `DefaultDisplayBitAsNumber`. If true, bit columns will be displayed as 1 or 0. If false, they'll be displayed as true or false. The default value is true, to keep parity with SSMS behavior.
Enables us to solve https://github.com/Microsoft/vscode-mssql/issues/690 and https://github.com/Microsoft/vscode-mssql/issues/513
Change logic for database change so OnInfoMessage does not get called until after the query execution is complete. This prevents multiple calls to open a single DataReader.
Adding new methods for executing queries from other services (such as the upcoming edit data service). The code is written to avoid duplicating logic by using lambdas to perform custom logic.
Additionally, the service host protocol has been slightly modified to split the IMessageSender into IEventSender and IRequestSender. This allows us to use either a ServiceHost or any RequestContext<T> to send events. It becomes very convenient to use another service's request context to send the events for query execution.
**Breaking Change**: This removes the messages property for query dispose results and instead elects to use error for any errors encountered during query disposal. A result is only used when something is successful.
* Splitting IMessageSender into IEventSender and IRequestSender
* Adding inter-service method for executing queries
* Adding inter-service method for disposing of a query
* Adding null checking for the success/error handlers
This is a small API addition that allows us to execute queries directly as strings. This will make it easier to execute queries outside the confines of a workspace like VS Code.
* Refactor execution requests and events are now named less redundantly and moved into a separate namespace for organization. This is the bulk of the changes.
* QueryExecuteBatchNotification -> ExecuteRequests/BatchEvents
* QueryExecuteMessageNotification -> ExecuteRequests/MessageEvent
* QueryExecuteCompleteNotification -> ExecuteRequests/QueryCompleteEvent
* QueryExecuteResultSetCompleteNotification -> ExecuteRequests/ResultSetEvents
* QueryExecuteSubsetRequest -> SubsetRequest.cs
* Creating an inheritance pattern where
* `ExecuteRequestParamsBase` has execution options and ID for a query execute request
* `ExecuteDocumentSelectionParams` inherits from `ExecuteRequestParamsBase` and provides a document selection
* `ExecuteStringParams` inherits from `ExecuteRequestParamsBase` and provides the query text
* Adding a helper method to get SQL text based on request type
* Through the AWESOME POWER OF POLYMORPHISM, we are able to create a request for executing straight SQL basically for free.
* **Breaking change:** query/execute => query/executeDocumentSelection to make it more obvious what is expected.
* Adding unit tests for the code that gets SQL text
* Refactoring of execute contracts into their own namespace
* Refactoring application
* Adding new request for executing queries as strings
* Adding forgotten string request
* Changing the logic for checking the request param types
* Removing redundant declarations
* Add CancelTokenKey for uniquely identifying cancelations of Connections associated with an OwnerUri and ConnectionType string.
* Update ConnectionInfo to use ConcurrentDictionary of DbConnection instances. Add wrapper functions for the ConcurrentDictionary.
* Refactor Connect and Disconnect in ConnectionService.
* Update ConnectionService: Handle multiple connections per ConnectionInfo. Handle cancelation tokens uniquely identified with CancelTokenKey. Add GetOrOpenConnection() for other services to request an existing or create a new DbConnection.
* Add ConnectionType.cs for ConnectionType strings.
* Add ConnectionType string to ConnectParams, ConnectionCompleteNotification, DisconnectParams.
* Update Query ExecuteInternal to use the dedicated query connection and GetOrOpenConnection().
* Update test library to account for multiple connections in ConnectionInfo.
* Write tests ensuring multiple connections don’t create redundant data.
* Write tests ensuring database changes affect all connections of a given ConnectionInfo.
* Write tests for TRANSACTION statements and temp tables.
* Fix exception due to repeat disposal of stream
* Fix off by one index for saving selection
* Fix to overwrite file if file already exists
* Fix test for Json save selection
* Add Json Formatting
* experimental showplan implementation (tools side only)
* fix for redundant massages from showplan executions
* moved showplan batches to new variables to make it less confusing
* returns showplan as part of batch summary with in each result summary
* cleaned up showplan resultsets
* cleaning up code and making showplan var optional
* changes all var names to showplan
* adding estimated support
* small fixes
* updatin var names and adding EPOptions struct
* adding ssms execution plan logic based on server types
* adding special actions logic
* removing redundant name changes
* execution plan query handler added
* cleaning up functions and data structures
* seperated special actions into its own class
* cleaning up special actions
* cleaning up code
* small new line fixes
* commenting out pre-yukon code
* removing all pre yukon code
* last yukon code commented out
* fixes broken tests
* adding related unit tests; integration tests incoming
* finishing tests and cleaning up code
* semantic changes
* cleaning up semantics
* changes and test fixes, also adding new exceptions into RS
* fixing special actions and cleaning up request logic
* fixing comment to trigger new build
* triggering another build
* fixed up specialaction and added tests for it
This change is a reworking of the way that messages are sent to clients from the service layer. It is also a reworking of the protocol to ensure that all formulations of query send back events to the client in a deterministic ordering. To support the first change:
* Added a new event that will be sent when a message is generated
* Messages now indicate which Batch (if any) generated them
* Messages now indicate if they were error level
* Removed message storage in Batch objects and BatchSummary objects
* Batch objects no longer have error state
It's an overhaul of the Save As mechanism to utilize the file reader/writer classes to better align with the patterns laid out by the rest of the query execution. Why make this change? This change makes our code base more uniform and adherent to the patterns/paradigms we've set up. This change also helps with the encapsulation of the classes to "separate the concerns" of each component of the save as function.
* Replumbing the save as execution to pass the call down the query stack as QueryExecutionService->Query->Batch->ResultSet
* Each layer performs it's own parameter checking
* QueryExecutionService checks if the query exists
* Query checks if the batch exists
* Batch checks if the result set exists
* ResultSet checks if the row counts are valid and if the result set has been executed
* Success/Failure delegates are passed down the chain as well
* Determination of whether a save request is a "selection" moved to the SaveResultsRequest class to eliminate duplication of code and creation of utility classes
* Making the IFileStream* classes more generic
* Removing the requirements of max characters to store from the GetWriter method, and moving it into the constructor for the temporary buffer writer - the values have been moved to the settings and given defaults
* Removing the individual type writers from IFileStreamWriter
* Removing the individual type writers from IFIleStreamReader
* Adding a new overload for WriteRow to IFileStreamWriter that will write out data, given a row's worth of data and the list of columns
* Creating a new IFileStreamFactory that creates a reader/writer pair for reading from the temporary files and writing to CSV files
* Creating a new IFileStreamFactory that creates a reader/writer pair for reading from the temporary files and writing to JSON files
* Dramatically simplified the CSV encoding functionality
* Removed duplicated logic for saving in different types and condensed down to a single chain that only differs based on what type of factory is provided
* Removing the logic for managing the list of save as tasks, since the ResultSet now performs the actual saving work, there's no real need to expose the internals of the ResultSet
* Adding new strings to the sr.strings file for save as error messages
* Completely rewriting the unit tests for the save as mechanism. Very fine grained unit tests now that should cover majority of cases (aside from race conditions)
* Refactoring maxchars params into settings and out of file stream factory
* Removing write*/read* methods from file stream readers/writers
* Migrating the CSV save as to the resultset
* Tweaks to unit testing to eliminate writing files to disk
* WIP, moving to a base class for save results writers
* Everything is wired up and compiles
* Adding unit tests for CSV encoding
* Adding unit tests for CSV and Json writers
* Adding tests to the result set for saving
* Refactor to throw exceptions on errors instead of calling failure handler
* Unit tests for batch/query argument in range
* Unit tests
* Adding service integration unit tests
* Final polish, copyright notices, etc
* Adding NULL logic
* Fixing issue of unicode to utf8
* Fixing issues as per @kburtram code review comments
* Adding files that got broken?
This basically is a replacement to the fix for Adding Milliseconds to DateTime fields. I didn't take into consideration that `DATE` columns would report as DateTime type. Date columns have a numeric scale set to 255, leading to the formatting string for the date time to include 255 millisecond places, which is invalid.
This change also reverses the change to store DateTime precision in the buffer file. Instead, the column metadata is now used when deserializing the rows from the db. `DATETIME` and `DATETIME2` columns are differentiated by their numeric scale while `DATE` columns are differentiated by their datatype name field.
More unit tests were added. Additionally, this fixes an unreported bug that `DATE` columns were being displayed with times, which is incorrect.
* Revert "Adding Milliseconds to DateTime fields (#173)"
This reverts commit 431dfa4156.
* Reworking the reader to use the column metadata for date types
* DbColumn -> DbColumnWrapper
* Fina tweaks to support DATETIME2(0)
* Removing the unused arguments
This change is a refactor of the DbColumnWrapper class. This fixes a huge oversight/gotcha in the DbColumnWrapper where treating the DbColumnWrapper as a DbColumn (ie, polymorphism) would result in almost all the fields of the column being null. This is a highly unexpected behavior. This change fixes that.
* DbColumn and ReliableConnection tests
* More retry connection tests
* More tests
* Fix broken peek definition integration tests
* Fix test bug
* Add a couple batch tests
* Add some more tests
* More tests for code coverage.
* Validation and Diagnostic tests
* A few more tests
* A few mote test changes.
* Update file path tests to run on Windows only
This is a slightly larger change than anticipated due the difference between `DATETIME`, `DATETIME2`, and `DateTime`. The `DATETIME` type always uses 3 decimal points of a second, while the `DATETIME2` type has 7 (although since `DATETIME2(7)` is default in SSMS suggesting that it is a variable precision type). Regardless of the db type, the engine returns `DateTime` C# type. The db types are only made visible via the column info, via the numeric precision and numeric scale. My findings were as such:
`DATETIME `: Precision = 23, Scale = 3
`DATETIME2`: Precision = 255, Scale = 7
The scale corresponds neatly with the number of second decimal points to show. The buffer file writer was modified to store both the scale and the number of ticks. Then the buffer file reader was modified to read in the precision and the number of ticks and generate the ToString version of the DateTime to add "f" as many times as there is scale, which corresponds to milliseconds.
* Code for writing milliseconds of datetime/datetime2 columns
* Adding unit tests
* Fixing potential bug with datetime2(0)
This change is part of the progressive results code. It will submit a notification from the service layer to indicate when execution of a batch has completed. This notification will contain the selection for batch, execution start time, and its ID. This will enable the extension to produce a header for the batch before the batch completes, in order to make it more clear to the user that execution is going on.
* Adding new event for batch start
* Unit tests
* Fixing comments as per @kevcunnane
The main change in this pull request is to add a new event that will be fired upon completion of a resultset but before the completion of a batch. This event will only fire if a resultset is available and generated.
Changes:
* ConnectionService - Slight changes to enable mocking, cleanup
* Batch - Moving summary generation into ResultSet class, adding generation of ordinals for resultset and locking of result set list (which needs further refinement, but would be outside scope of this change)
* Adding new event and associated parameters for completion of a resultset. Params return the resultset summary
* Adding logic for assigning the event a handler in the query execution service
* Adding unit tests for testing the new event /making sure the existing tests work
* Refactoring some private properties into member variables
* Refactor to remove SectionData class in favor of BufferRange
* Adding callback for batch completion that will let the extension know that a batch has completed execution
* Refactoring to make progressive results work as per async query execution
* Allowing retrieval of batch results while query is in progress
* reverting global.json, whoops
* Adding a few missing comments, and fixing a couple code style bugs
* Using SelectionData everywhere again
* One more missing comment
* Adding new notification type for result set completion
* Plumbing event for result set completion
* Unit tests for result set events
This includes a fairly substantial change to create a mock of the
ConnectionService and to create separate memorystream storage arrays. It
preserves more correct behavior with a integration test, fixes an issue
where the test db reader will return n-1 rows because the Reliable
Connection Helper steals a record.
* Adding locking to ResultSets for thread safety
* Adding/fixing unit tests
* Adding batch ID to result set summary
This is a fairly large set of changes to the unit tests that help isolate the effectiveness of the unit tests.
* Unit tests for query execution have been split into separate files for different classes.
* Unit tests have been added for the ResultSet class which previously did not have tests
* The InMemoryStreamWrapper has been improved to share memory, creating a simulated filesystem
* Creating a mock ConnectionService to decrease noisy exceptions and prevent "row stealing". Unfortunately this lowers code coverage. However, since the tests that touched the connection service were not really testing it, this helps keep us honest. But it will require adding more unit tests for connection service.
* Standardizing the await mechanism for query execution
* Cleaning up the mechanism for getting WorkspaceService mocks and mock FileStreamFactories
* Refactor the query execution tests into their own files
* Removing tests from ExecuteTests.cs that were moved to separate files
* Adding tests for ResultSet class
* Adding test for the FOR XML/JSON component of the resultset class
* Setting up shared storage between file stream readers/writers
* Standardizing on Workspace mocking, awaiting execution completion
* Adding comment for ResultSet class
The main feature of this pull request is a new callback that's added to the query class that is called when a batch has completed execution and retrieval of results. This callback will send an event to the extension with the batch summary information. After that, the extension can submit subset requests for the resultsets of the batch.
Other smaller changes in this pull request:
Refactor to assign a batch a id when its created instead of when returning the list of batch summaries
Passing the SelectionData around instead of extracting the values for it
Moving creation of BatchSummary into the Batch class
Retrieval of results is now permitted even if the entire query has not completed, as long as the batch requested has completed.
Also note, this does not break the protocol. It adds a new event that a queryRunner can listen to, but it doesn't require it to be listened to.
* Refactor to remove SectionData class in favor of BufferRange
* Adding callback for batch completion that will let the extension know that a batch has completed execution
* Refactoring to make progressive results work as per async query execution
* Allowing retrieval of batch results while query is in progress
* reverting global.json, whoops
* Adding a few missing comments, and fixing a couple code style bugs
* Using SelectionData everywhere again
* One more missing comment
* Make save results asynchronous
* Prevent write share of file
* Lock objects in stages
* Create Save result objects
* refactor and write rows in batches
* CHange batchSize from test value
* Remove await in handler
* Removing the file reader as a member of the resultset
* Change Dispose to wait for save
* Change concurrentBag
* PascalCase variables
* Modify function signature and tests
* Safe file methods
* refactor ResultSets to Ilist and remove ToList
* Change dictionary key and prevent add to saveTasks during dispose
* Simplify row concatenation
* Fix prevent add
* Fix prevent add
* Add methods to expose saveTasks and isBeingDisposed
After much thinking, this change brings the message behavior into line with SSMS, including if the NOCOUNT is set. This is a somewhat non-obvious solution, but the StatementCompleted event handler will only be fired if there is a record count to return. We'll add the message for number of records if the StatementCompleted event is fired, otherwise we won't add any messages while processing the resultsets of the batch. If any messages are returned from the server, we'll capture those. Then, if at the end of the batch, we haven't collected any messages from StatementCompleted events or server messages, then we'll add the "completed successfully" message.
This matches behavior of SSMS that will only emit a "completed successfully" message if there were no other messages for the batch.
* Solution to issue, some unit tests needed to be tweaked
* Comments for the event handler
Moving some logic around such that when a query is cancelled, it isn't thrown away, allowing reading of the partial results, a la SSMS.
Adding a configure await to fix a tenacious bug causing query cancellations to hang for 20s or more.
Capturing sql errors for user cancellation, to return all user cancelation scenarios using the same messages. (ie, cancelling during ExecuteReaderAsync will yield a error from the server whereas cancelling during ReadAsync throws a TaskCancelledException
No changes to protocol, just implementation changes.
* Test of try/finally
* Fixed issue where resultsets are unreadable after query cancellation
* Fix for await/async issue
The two main changes in this pull request:
Launching query execution as an asynchronous task that performs a callback upon completion or failure of a query. (Which also sets us up for callbacks progressive results)
Moving away from using the Result of a query execution to return an error. Instead we'll use an error event to return an error
Additionally, some nice refactoring and cleaning up of the unit tests to take advantage of the cool RequestContext mock tooling by @kevcunnane
* Initial commit of refactor to run execution truely asynchronously
* Moving the storage of the task into Query class
Callbacks for completion of a query and failure of a query are setup as
events in the Query class. This actually sets us up for a very nice
framework for adding batch and resultset completion callbacks.
However, this also exposes a problem with cancelling queries and returning
errors -- we don't properly handle errors during execution of a query
(aside from DB errors).
* Wrapping things up in order to submit for code review
* Adding fixes as per comments
Fixing a bug where in various situations, the files used for temporary storage of query results would be leftover. In particular, the following changes were made:
* When the dispose query request is submitted, the corresponding query is now disposed in addition from being removed from the list of active queries
* When a query is cancelled, it is disposed after it is cancelled
* If a query already exists for a given ownerURI, the existing query is disposed before creating a new query
* All queries are disposed when the query execution service is disposed (ie, at shutdown of the service)
A unit test to verify the action of the dispose method for a ResultSet was added.
* Ensuring queries are disposed
Adding logic to dispose any queries when:
* URI that already has a query executes another query
* A request to dispose a query is submitted
* A request to cancel a query is submitted
* Small tweaks for cleanup of query execution service
* added support for timestamps
* fixed tests
* Moved message class to own file; added 'z' to end of date strings
* added default time constructor
* removed unnecessary z
* added time string format info in comment
* changed from utc time to using local time
