9.10. Writing a Custom Scan Provider
PostgreSQL supports a set of experimental facilities which are intended to allow extension modules to add new scan types to the system. Unlike a linkend=»fdwhandler»>foreign data wrapper, which is only responsible for knowing how to scan its own foreign tables, a custom scan provider can provide an alternative method of scanning any relation in the system. Typically, the motivation for writing a custom scan provider will be to allow the use of some optimization not supported by the core system, such as caching or some form of hardware acceleration. This chapter outlines how to write a new custom scan provider.
Implementing a new type of custom scan is a three-step process. First, during planning, it is necessary to generate access paths representing a scan using the proposed strategy. Second, if one of those access paths is selected by the planner as the optimal strategy for scanning a particular relation, the access path must be converted to a plan. Finally, it must be possible to execute the plan and generate the same results that would have been generated for any other access path targeting the same relation.
9.10.1. Creating Custom Scan Paths
A custom scan provider will typically add paths for a base relation by setting the following hook, which is called after the core code has generated all the access paths it can for the relation (except for Gather paths, which are made after this call so that they can use partial paths added by the hook):
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
RelOptInfo *rel,
Index rti,
RangeTblEntry *rte);
extern PGDLLIMPORT set_rel_pathlist_hook_type set_rel_pathlist_hook;
Although this hook function can be used to examine, modify, or remove paths generated by the core system, a custom scan provider will typically confine itself to generating CustomPath objects and adding them to rel using add_path. The custom scan provider is responsible for initializing the CustomPath object, which is declared like this:
typedef struct CustomPath
{
Path path;
uint32 flags;
List *custom_paths;
List *custom_private;
const CustomPathMethods *methods;
} CustomPath;
path must be initialized as for any other path, including the row-count estimate, start and total cost, and sort ordering provided by this path. flags is a bit mask, which specifies whether the scan provider can support certain optional capabilities. flags should include CUSTOMPATH_SUPPORT_BACKWARD_SCAN if the custom path can support a backward scan, CUSTOMPATH_SUPPORT_MARK_RESTORE if it can support mark and restore, and CUSTOMPATH_SUPPORT_PROJECTION if it can perform projections. (If CUSTOMPATH_SUPPORT_PROJECTION is not set, the scan node will only be asked to produce Vars of the scanned relation; while if that flag is set, the scan node must be able to evaluate scalar expressions over these Vars.) An optional custom_paths is a list of Path nodes used by this custom-path node; these will be transformed into Plan nodes by planner. custom_private can be used to store the custom path’s private data. Private data should be stored in a form that can be handled by nodeToString, so that debugging routines that attempt to print the custom path will work as designed. methods must point to a (usually statically allocated) object implementing the required custom path methods, of which there is currently only one.
A custom scan provider can also provide join paths. Just as for base relations, such a path must produce the same output as would normally be produced by the join it replaces. To do this, the join provider should set the following hook, and then within the hook function, create CustomPath path(s) for the join relation.
typedef void (*set_join_pathlist_hook_type) (PlannerInfo *root,
RelOptInfo *joinrel,
RelOptInfo *outerrel,
RelOptInfo *innerrel,
JoinType jointype,
JoinPathExtraData *extra);
extern PGDLLIMPORT set_join_pathlist_hook_type set_join_pathlist_hook;
This hook will be invoked repeatedly for the same join relation, with different combinations of inner and outer relations; it is the responsibility of the hook to minimize duplicated work.
9.10.1.1. Custom Scan Path Callbacks
Plan *(*PlanCustomPath) (PlannerInfo *root,
RelOptInfo *rel,
CustomPath *best_path,
List *tlist,
List *clauses,
List *custom_plans);
Convert a custom path to a finished plan. The return value will generally
be a CustomScan object, which the callback must allocate and initialize. See custom-scan-plan for more details.
9.10.2. Creating Custom Scan Plans
A custom scan is represented in a finished plan tree using the following structure:
typedef struct CustomScan
{
Scan scan;
uint32 flags;
List *custom_plans;
List *custom_exprs;
List *custom_private;
List *custom_scan_tlist;
Bitmapset *custom_relids;
const CustomScanMethods *methods;
} CustomScan;
scan must be initialized as for any other scan, including estimated costs, target lists, qualifications, and so on. flags is a bit mask with the same meaning as in CustomPath. custom_plans can be used to store child Plan nodes. custom_exprs should be used to store expression trees that will need to be fixed up by setrefs.c and subselect.c, while custom_private should be used to store other private data that is only used by the custom scan provider itself. custom_scan_tlist can be NIL when scanning a base relation, indicating that the custom scan returns scan tuples that match the base relation’s row type. Otherwise it is a target list describing the actual scan tuples. custom_scan_tlist must be provided for joins, and could be provided for scans if the custom scan provider can compute some non-Var expressions. custom_relids is set by the core code to the set of relations (range table indexes) that this scan node handles; except when this scan is replacing a join, it will have only one member. methods must point to a (usually statically allocated) object implementing the required custom scan methods, which are further detailed below.
When a CustomScan scans a single relation, scan.scanrelid must be the range table index of the table to be scanned. When it replaces a join, scan.scanrelid should be zero.
Plan trees must be able to be duplicated using copyObject, so all the data stored within the custom fields must consist of nodes that that function can handle. Furthermore, custom scan providers cannot substitute a larger structure that embeds a CustomScan for the structure itself, as would be possible for a CustomPath or CustomScanState.
9.10.2.1. Custom Scan Plan Callbacks
Node *(*CreateCustomScanState) (CustomScan *cscan);
Allocate a **CustomScanState** for this
CustomScan. The actual allocation will often be larger than required for an ordinary CustomScanState, because many providers will wish to embed that as the first field of a larger structure. The value returned must have the node tag and methods set appropriately, but other fields should be left as zeroes at this stage; after ExecInitCustomScan performs basic initialization, the BeginCustomScan callback will be invoked to give the custom scan provider a chance to do whatever else is needed.
9.10.3. Executing Custom Scans
When a CustomScan is executed, its execution state is represented by a CustomScanState, which is declared as follows:
typedef struct CustomScanState
{
ScanState ss;
uint32 flags;
const CustomExecMethods *methods;
} CustomScanState;
ss is initialized as for any other scan state, except that if the scan is for a join rather than a base relation, ss.ss_currentRelation is left NULL. flags is a bit mask with the same meaning as in CustomPath and CustomScan. methods must point to a (usually statically allocated) object implementing the required custom scan state methods, which are further detailed below. Typically, a CustomScanState, which need not support copyObject, will actually be a larger structure embedding the above as its first member.
9.10.3.1. Custom Scan Execution Callbacks
void (*BeginCustomScan) (CustomScanState *node,
EState *estate,
int eflags);
Complete initialization of the supplied **CustomScanState**.
Standard fields have been initialized by ExecInitCustomScan, but any private fields should be initialized here.
TupleTableSlot *(*ExecCustomScan) (CustomScanState *node);
Fetch the next scan tuple. If any tuples remain, it should fill
ps_ResultTupleSlot with the next tuple in the current scan direction, and then return the tuple slot. If not, NULL or an empty slot should be returned.
void (*EndCustomScan) (CustomScanState *node);
Clean up any private data associated with the **CustomScanState**.
This method is required, but it does not need to do anything if there is no associated data or it will be cleaned up automatically.
void (*ReScanCustomScan) (CustomScanState *node);
Rewind the current scan to the beginning and prepare to rescan the
relation.
void (*MarkPosCustomScan) (CustomScanState *node);
Save the current scan position so that it can subsequently be restored
by the RestrPosCustomScan callback. This callback is optional, and need only be supplied if the CUSTOMPATH_SUPPORT_MARK_RESTORE flag is set.
void (*RestrPosCustomScan) (CustomScanState *node);
Restore the previous scan position as saved by the
MarkPosCustomScan callback. This callback is optional, and need only be supplied if the CUSTOMPATH_SUPPORT_MARK_RESTORE flag is set.
Size (*EstimateDSMCustomScan) (CustomScanState *node,
ParallelContext *pcxt);
Estimate the amount of dynamic shared memory that will be required
for parallel operation. This may be higher than the amount that will actually be used, but it must not be lower. The return value is in bytes. This callback is optional, and need only be supplied if this custom scan provider supports parallel execution.
void (*InitializeDSMCustomScan) (CustomScanState *node,
ParallelContext *pcxt,
void *coordinate);
Initialize the dynamic shared memory that will be required for parallel
operation. coordinate points to a shared memory area of size equal to the return value of EstimateDSMCustomScan. This callback is optional, and need only be supplied if this custom scan provider supports parallel execution.
void (*ReInitializeDSMCustomScan) (CustomScanState *node,
ParallelContext *pcxt,
void *coordinate);
Re-initialize the dynamic shared memory required for parallel operation
when the custom-scan plan node is about to be re-scanned. This callback is optional, and need only be supplied if this custom scan provider supports parallel execution. Recommended practice is that this callback reset only shared state, while the ReScanCustomScan callback resets only local state. Currently, this callback will be called before ReScanCustomScan, but it’s best not to rely on that ordering.
void (*InitializeWorkerCustomScan) (CustomScanState *node,
shm_toc *toc,
void *coordinate);
Initialize a parallel worker's local state based on the shared state
set up by the leader during InitializeDSMCustomScan. This callback is optional, and need only be supplied if this custom scan provider supports parallel execution.
void (*ShutdownCustomScan) (CustomScanState *node);
Release resources when it is anticipated the node will not be executed
to completion. This is not called in all cases; sometimes, EndCustomScan may be called without this function having been called first. Since the DSM segment used by parallel query is destroyed just after this callback is invoked, custom scan providers that wish to take some action before the DSM segment goes away should implement this method.
void (*ExplainCustomScan) (CustomScanState *node,
List *ancestors,
ExplainState *es);
Output additional information for **EXPLAIN** of a custom-scan
plan node. This callback is optional. Common data stored in the ScanState, such as the target list and scan relation, will be shown even without this callback, but the callback allows the display of additional, private state.