mirror of
https://github.com/postgres/postgres.git
synced 2025-04-22 23:02:54 +03:00
src/backend/executor/README The Postgres Executor ===================== The executor processes a tree of "plan nodes". The plan tree is essentially a demand-pull pipeline of tuple processing operations. Each node, when called, will produce the next tuple in its output sequence, or NULL if no more tuples are available. If the node is not a primitive relation-scanning node, it will have child node(s) that it calls in turn to obtain input tuples. Refinements on this basic model include: * Choice of scan direction (forwards or backwards). Caution: this is not currently well-supported. It works for primitive scan nodes, but not very well for joins, aggregates, etc. * Rescan command to reset a node and make it generate its output sequence over again. * Parameters that can alter a node's results. After adjusting a parameter, the rescan command must be applied to that node and all nodes above it. There is a moderately intelligent scheme to avoid rescanning nodes unnecessarily (for example, Sort does not rescan its input if no parameters of the input have changed, since it can just reread its stored sorted data). For a SELECT, it is only necessary to deliver the top-level result tuples to the client. For INSERT/UPDATE/DELETE, the actual table modification operations happen in a top-level ModifyTable plan node. If the query includes a RETURNING clause, the ModifyTable node delivers the computed RETURNING rows as output, otherwise it returns nothing. Handling INSERT is pretty straightforward: the tuples returned from the plan tree below ModifyTable are inserted into the correct result relation. For UPDATE, the plan tree returns the computed tuples to be updated, plus a "junk" (hidden) CTID column identifying which table row is to be replaced by each one. For DELETE, the plan tree need only deliver a CTID column, and the ModifyTable node visits each of those rows and marks the row deleted. XXX a great deal more documentation needs to be written here... Plan Trees and State Trees -------------------------- The plan tree delivered by the planner contains a tree of Plan nodes (struct types derived from struct Plan). Each Plan node may have expression trees associated with it, to represent its target list, qualification conditions, etc. During executor startup we build a parallel tree of identical structure containing executor state nodes --- every plan and expression node type has a corresponding executor state node type. Each node in the state tree has a pointer to its corresponding node in the plan tree, plus executor state data as needed to implement that node type. This arrangement allows the plan tree to be completely read-only as far as the executor is concerned: all data that is modified during execution is in the state tree. Read-only plan trees make life much simpler for plan caching and reuse. Altogether there are four classes of nodes used in these trees: Plan nodes, their corresponding PlanState nodes, Expr nodes, and their corresponding ExprState nodes. (Actually, there are also List nodes, which are used as "glue" in all four kinds of tree.) Memory Management ----------------- A "per query" memory context is created during CreateExecutorState(); all storage allocated during an executor invocation is allocated in that context or a child context. This allows easy reclamation of storage during executor shutdown --- rather than messing with retail pfree's and probable storage leaks, we just destroy the memory context. In particular, the plan state trees and expression state trees described in the previous section are allocated in the per-query memory context. To avoid intra-query memory leaks, most processing while a query runs is done in "per tuple" memory contexts, which are so-called because they are typically reset to empty once per tuple. Per-tuple contexts are usually associated with ExprContexts, and commonly each PlanState node has its own ExprContext to evaluate its qual and targetlist expressions in. Query Processing Control Flow ----------------------------- This is a sketch of control flow for full query processing: CreateQueryDesc ExecutorStart CreateExecutorState creates per-query context switch to per-query context to run ExecInitNode ExecInitNode --- recursively scans plan tree CreateExprContext creates per-tuple context ExecInitExpr AfterTriggerBeginQuery ExecutorRun ExecProcNode --- recursively called in per-query context ExecEvalExpr --- called in per-tuple context ResetExprContext --- to free memory ExecutorFinish ExecPostprocessPlan --- run any unfinished ModifyTable nodes AfterTriggerEndQuery ExecutorEnd ExecEndNode --- recursively releases resources FreeExecutorState frees per-query context and child contexts FreeQueryDesc Per above comments, it's not really critical for ExecEndNode to free any memory; it'll all go away in FreeExecutorState anyway. However, we do need to be careful to close relations, drop buffer pins, etc, so we do need to scan the plan state tree to find these sorts of resources. The executor can also be used to evaluate simple expressions without any Plan tree ("simple" meaning "no aggregates and no sub-selects", though such might be hidden inside function calls). This case has a flow of control like CreateExecutorState creates per-query context CreateExprContext -- or use GetPerTupleExprContext(estate) creates per-tuple context ExecPrepareExpr temporarily switch to per-query context run the expression through expression_planner ExecInitExpr Repeatedly do: ExecEvalExprSwitchContext ExecEvalExpr --- called in per-tuple context ResetExprContext --- to free memory FreeExecutorState frees per-query context, as well as ExprContext (a separate FreeExprContext call is not necessary) EvalPlanQual (READ COMMITTED Update Checking) --------------------------------------------- For simple SELECTs, the executor need only pay attention to tuples that are valid according to the snapshot seen by the current transaction (ie, they were inserted by a previously committed transaction, and not deleted by any previously committed transaction). However, for UPDATE and DELETE it is not cool to modify or delete a tuple that's been modified by an open or concurrently-committed transaction. If we are running in SERIALIZABLE isolation level then we just raise an error when this condition is seen to occur. In READ COMMITTED isolation level, we must work a lot harder. The basic idea in READ COMMITTED mode is to take the modified tuple committed by the concurrent transaction (after waiting for it to commit, if need be) and re-evaluate the query qualifications to see if it would still meet the quals. If so, we regenerate the updated tuple (if we are doing an UPDATE) from the modified tuple, and finally update/delete the modified tuple. SELECT FOR UPDATE/SHARE behaves similarly, except that its action is just to lock the modified tuple and return results based on that version of the tuple. To implement this checking, we actually re-run the query from scratch for each modified tuple (or set of tuples, for SELECT FOR UPDATE), with the relation scan nodes tweaked to return only the current tuples --- either the original ones, or the updated (and now locked) versions of the modified tuple(s). If this query returns a tuple, then the modified tuple(s) pass the quals (and the query output is the suitably modified update tuple, if we're doing UPDATE). If no tuple is returned, then the modified tuple(s) fail the quals, so we ignore the current result tuple and continue the original query. In UPDATE/DELETE, only the target relation needs to be handled this way. In SELECT FOR UPDATE, there may be multiple relations flagged FOR UPDATE, so we obtain lock on the current tuple version in each such relation before executing the recheck. It is also possible that there are relations in the query that are not to be locked (they are neither the UPDATE/DELETE target nor specified to be locked in SELECT FOR UPDATE/SHARE). When re-running the test query we want to use the same rows from these relations that were joined to the locked rows. For ordinary relations this can be implemented relatively cheaply by including the row TID in the join outputs and re-fetching that TID. (The re-fetch is expensive, but we're trying to optimize the normal case where no re-test is needed.) We have also to consider non-table relations, such as a ValuesScan or FunctionScan. For these, since there is no equivalent of TID, the only practical solution seems to be to include the entire row value in the join output row. We disallow set-returning functions in the targetlist of SELECT FOR UPDATE, so as to ensure that at most one tuple can be returned for any particular set of scan tuples. Otherwise we'd get duplicates due to the original query returning the same set of scan tuples multiple times. (Note: there is no explicit prohibition on SRFs in UPDATE, but the net effect will be that only the first result row of an SRF counts, because all subsequent rows will result in attempts to re-update an already updated target row. This is historical behavior and seems not worth changing.)