Refactor txn context management in session

[lcwangchao]

Introduction

We have some problems for the current transaction context implementation:

No single truth for txn context

Though we have sessionVars.TxnCtx to store the startTS and InfoSchema for current txn, we still need to obtain them from other places in some cases. For example, when we are using stale read in implicit transaction to read data, TxnCtx.InfoSchema and TxnCtx.StartTS is not set with right values. Instead, the real values are computed in session and passed to planner and executors by method arguments.

It seems that sessionctx.Context.GetInfoSchema provides the "current" infoschema. But it is not true, it provides the latest infoschema in implicit transaction. This behavior misleads developers a lot.

The codes maintaining txn context in different scenarios are serious coupling

We have several scenarios (i.e. RR/RC isolation, stale/historical read ...) providing txn context. These codes are mixed up not have not been modularized. So it is hard to read these codes.

Design

To solve the above problems. We'll introduce a new interface TxnManager which provides some methods to manage transaction context in session. The interface is like this:

pacakge sessiontxn

type TxnManager interface {
  GetTxnInfoSchema(ctx context.Context) (infoschema.InfoSchema, error)
  GetStmtReadTS(ctx context.Context) (uint64, error)
  ...
}

The get method, for example, GetTxnInfoSchema should be a single truth to return the real txn infoschema considering all scenarios.

For maintaining the context for different scenarios, we can also introduce a new interface TxnContextProvider :

type TxnContextProvider interface {
  Initialize(sctx sessionctx.Context) error;
  OnStartStmt() error;
  Txn(active bool) (kv.Transaction, error)
  GetTxnInfoSchema(ctx context.Context) (infoschema.InfoSchema, error)
  GetStmtReadTS(ctx context.Context) (uint64, error)
  ...
}

It provides two kind of methods. The first methods provide some lifecycle hooks, they should be called by TxnManager at the corresponding stage. The second methods provide context for their own scenario. After this, we can have RRIsolationContextProvider , RCIsolationContextProvider , StaleReadContextProvider ...

TxnManager should provide some methods to get/set the current provider:

type TxnContextProvider interface {
  SetContextProvider(provider *TxnContextProvider) error
  GetContextProvider() TxnContextProvider
  ...
}

We can obtain TxnManager from session like this:

txnManager := sessiontxn.GetTxnManager(sctx)

GetTxnManager is defined as a factory method:

func GetTxnManager(sctx sessionctx.Context) TxnManager {
  // Get txnManager from session
}

Refactor path

1. At first step of refactoring, we can implement the "read" methods of the TxnManager which only "wraps" the current implements. At this time, SimpleTxnContextProvider can be provided to do the wrap works. We should also add some unit tests to ensure that the context get from TxnManager and the old interface equals.

2. After step 1, the codes are still using the old interface to get the txn context. Then we should migrate these codes to reading the new interface for TxnManager . If we did well in step1, this step should be safe.

3. After all interfaces are migrated to TxnManager . We can introduce more context providers to replace SimpleTxnContextProvider for each scenario. For example, we can introduce StaleReadTxnContextProvider to provide stale read context.

Tasks

• Introduce TxnManager and implement GetTxnInfoSchema() in it. #30695
• Use TxnManager.GetTxnInfoSchema() to get the txn infoschema #30933
• Add GetReadTS and GetForUpdateTS for TxnManager #33014
• Add EnterNewTxn and OnStmtStart to manage txn lifecycle #34390
• Add error handling methods for TxnManager #34533
• Stale Read
  • A draft for stale read refactor *: (DRAFT) Refactor stale read code #32325
  • Provide staleReadProcessor to process stale read #32697
  • refactor execute prepared statement for stale read #32938
  • Move the stale read context to the txn provider #33832
  • Remove some useless stale read code left after refactored #35848
• Refactor Txn in RC isolation #34088
  • *: refactor rc read #33995
  • Introduce txn context provider for pessimistic RC isolation level. #34746
  • RCCheckTS should be managed by RC transaction context provider. #35497
• Refactor RR in pessimistic mode
  • Implement provider for RR #35129
• Refactor Serializable
  • Introduce txn context provider for pessimistic Serializable isolation level. #35150
• Refactor RR in optimistic mode
  • Introduce OptimisticTxnContextProvider for optimistic txn #35130
• local txn
  • Decouple the term TxnScope from the Stale Read code #31772
• Misc
  • Remove legacy.SimpleTxnContextProvider #35666
  • Anomalies are detected by jepsen bank-multitable #35459
  • Unify the entrance of new transaction. #35705
  • Snapshot acquisition should be managed by transaction manager #35810
  • Make the code cleaner in session.ExecutePreparedStmt #35925
  • add checks when EnterNewTxnBeforeStmt #35959
  • move some methods in session_txn to internal pacakge #36033

[cfzjywxk]

/cc @youjiali1995 @sticnarf @you06 @MyonKeminta @ekexium @longfangsong