RowCountable
, StoreCostController
public class HeapCostController extends GenericCostController implements StoreCostController
It is likely that the implementation of StoreCostController will open the conglomerate and will leave the conglomerate open until the StoreCostController is closed. This represents a significant amount of work, so the caller if possible should attempt to open the StoreCostController once per unit of work and rather than close and reopen the controller. For instance if the optimizer needs to cost 2 different scans against a single conglomerate, it should use one instance of the StoreCostController.
The locking behavior of the implementation of a StoreCostController is undefined, it may or may not get locks on the underlying conglomerate. It may or may not hold locks until end of transaction. An optimal implementation will not get any locks on the underlying conglomerate, thus allowing concurrent access to the table by a executing query while another query is optimizing.
The StoreCostController gives 2 kinds of cost information
Modifier and Type | Field | Description |
---|---|---|
(package private) long |
num_pages |
Only lookup these estimates from raw store once.
|
(package private) long |
num_rows |
|
(package private) long |
page_size |
|
(package private) long |
row_size |
open_conglom
BASE_CACHED_ROW_FETCH_COST, BASE_GROUPSCAN_ROW_COST, BASE_HASHSCAN_ROW_FETCH_COST, BASE_NONGROUPSCAN_ROW_FETCH_COST, BASE_ROW_PER_BYTECOST, BASE_UNCACHED_ROW_FETCH_COST, STORECOST_CLUSTERED, STORECOST_SCAN_NORMAL, STORECOST_SCAN_SET
Constructor | Description |
---|---|
HeapCostController() |
Modifier and Type | Method | Description |
---|---|---|
double |
getFetchFromRowLocationCost(FormatableBitSet validColumns,
int access_type) |
Return the cost of calling ConglomerateController.fetch().
|
void |
getScanCost(int scan_type,
long row_count,
int group_size,
boolean forUpdate,
FormatableBitSet scanColumnList,
DataValueDescriptor[] template,
DataValueDescriptor[] startKeyValue,
int startSearchOperator,
DataValueDescriptor[] stopKeyValue,
int stopSearchOperator,
boolean reopen_scan,
int access_type,
StoreCostResult cost_result) |
Calculate the cost of a scan.
|
void |
init(OpenConglomerate open_conglom) |
Initialize the cost controller.
|
checkConsistency, close, debugConglomerate, getEstimatedRowCount, getInternalTablePropertySet, getOpenConglom, getRowPositionFromRowLocation, getSpaceInfo, getTableProperties, isKeyed, isTableLocked, newRowLocationTemplate, queueDeletePostCommitWork, setEstimatedRowCount
getFetchFromFullKeyCost
clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait
getEstimatedRowCount, setEstimatedRowCount
close, getFetchFromFullKeyCost, newRowLocationTemplate
long num_pages
long num_rows
long page_size
long row_size
public void init(OpenConglomerate open_conglom) throws StandardException
Let super.init() do it's work and then get the initial stats about the table from raw store.
init
in class GenericController
StandardException
- Standard exception policy.public double getFetchFromRowLocationCost(FormatableBitSet validColumns, int access_type) throws StandardException
Return the estimated cost of calling ConglomerateController.fetch() on the current conglomerate. This gives the cost of finding a record in the conglomerate given the exact RowLocation of the record in question.
The validColumns describes what kind of row is being fetched, ie. it may be cheaper to fetch a partial row than a complete row.
getFetchFromRowLocationCost
in interface StoreCostController
validColumns
- A description of which columns to return from
row on the page into "templateRow." templateRow,
and validColumns work together to
describe the row to be returned by the fetch -
see RowUtil for description of how these three
parameters work together to describe a fetched
"row".access_type
- Describe the type of access the query will be
performing to the ConglomerateController.
STORECOST_CLUSTERED - The location of one fetch
is likely clustered "close" to the next
fetch. For instance if the query plan were
to sort the RowLocations of a heap and then
use those RowLocations sequentially to
probe into the heap, then this flag should
be specified. If this flag is not set then
access to the table is assumed to be
random - ie. the type of access one gets
if you scan an index and probe each row
in turn into the base table is "random".StandardException
- Standard exception policy.RowUtil
public void getScanCost(int scan_type, long row_count, int group_size, boolean forUpdate, FormatableBitSet scanColumnList, DataValueDescriptor[] template, DataValueDescriptor[] startKeyValue, int startSearchOperator, DataValueDescriptor[] stopKeyValue, int stopSearchOperator, boolean reopen_scan, int access_type, StoreCostResult cost_result) throws StandardException
Cause this object to calculate the cost of performing the described scan. The interface is setup such that first a call is made to calcualteScanCost(), and then subsequent calls to accessor routines are made to get various pieces of information about the cost of the scan.
For the purposes of costing this routine is going to assume that a page will remain in cache between the time one next()/fetchNext() call and a subsequent next()/fetchNext() call is made within a scan.
The result of costing the scan is placed in the "cost_result". The cost of the scan is stored by calling cost_result.setEstimatedCost(cost). The estimated row count is stored by calling cost_result.setEstimatedRowCount(row_count).
The estimated cost of the scan assumes the caller will execute a fetchNext() loop for every row that qualifies between start and stop position. Note that this cost is different than execution a next(),fetch() loop; or if the scan is going to be terminated by client prior to reaching the stop condition.
The estimated number of rows returned from the scan assumes the caller will execute a fetchNext() loop for every row that qualifies between start and stop position.
getScanCost
in interface StoreCostController
scan_type
- The type of scan that will be executed. There
are currently 2 types:
STORECOST_SCAN_NORMAL - scans will be executed
using the standard next/fetch, where each fetch
can retrieve 1 or many rows (if fetchNextGroup()
interface is used).
STORECOST_SCAN_SET - The entire result set will
be retrieved using the the fetchSet() interface.row_count
- Estimated total row count of the table. The
current system tracks row counts in heaps better
than btree's (btree's have "rows" which are not
user rows - branch rows, control rows), so
if available the client should
pass in the base table's row count into this
routine to be used as the index's row count.
If the caller has no idea, pass in -1.group_size
- The number of rows to be returned by a single
fetch call for STORECOST_SCAN_NORMAL scans.forUpdate
- Should be true if the caller intends to update
through the scan.scanColumnList
- A description of which columns to return from
every fetch in the scan. template,
and scanColumnList work together
to describe the row to be returned by the scan -
see RowUtil for description of how these three
parameters work together to describe a "row".template
- A prototypical row which the scan may use to
maintain its position in the conglomerate. Not
all access method scan types will require this,
if they don't it's ok to pass in null.
In order to scan a conglomerate one must
allocate 2 separate "row" templates. The "row"
template passed into openScan is for the private
use of the scan itself, and no access to it
should be made by the caller while the scan is
still open. Because of this the scanner must
allocate another "row" template to hold the
values returned from fetch(). Note that this
template must be for the full row, whether a
partial row scan is being executed or not.startKeyValue
- An indexable row which holds a (partial) key
value which, in combination with the
startSearchOperator, defines the starting
position of the scan. If null, the starting
position of the scan is the first row of the
conglomerate. The startKeyValue must only
reference columns included in the scanColumnList.startSearchOperator
- an operator which defines how the startKeyValue
is to be searched for. If startSearchOperation
is ScanController.GE, the scan starts on the
first row which is greater than or equal to the
startKeyValue. If startSearchOperation is
ScanController.GT, the scan starts on the first
row whose key is greater than startKeyValue. The
startSearchOperation parameter is ignored if the
startKeyValue parameter is null.stopKeyValue
- An indexable row which holds a (partial) key
value which, in combination with the
stopSearchOperator, defines the ending position
of the scan. If null, the ending position of the
scan is the last row of the conglomerate. The
stopKeyValue must only reference columns included
in the scanColumnList.stopSearchOperator
- an operator which defines how the stopKeyValue
is used to determine the scan stopping position.
If stopSearchOperation is ScanController.GE, the
scan stops just before the first row which is
greater than or equal to the stopKeyValue. If
stopSearchOperation is ScanController.GT, the
scan stops just before the first row whose key
is greater than startKeyValue. The
stopSearchOperation parameter is ignored if the
stopKeyValue parameter is null.access_type
- Describe the type of access the query will be
performing to the ScanController.
STORECOST_CLUSTERED - The location of one scan
is likely clustered "close" to the previous
scan. For instance if the query plan were
to used repeated "reopenScan()'s" to probe
for the next key in an index, then this flag
should be be specified. If this flag is not
set then each scan will be costed independant
of any other predicted scan access.StandardException
- Standard exception policy.RowUtil
Apache Derby V10.14 Internals - Copyright © 2004,2018 The Apache Software Foundation. All Rights Reserved.