englefly commented on code in PR #69322: URL: https://github.com/apache/doris/pull/69322#discussion_r4335584466
########## fe/fe-core/src/main/java/org/apache/doris/nereids/stats/HboStructFreshness.java: ########## @@ -0,0 +1,682 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +package org.apache.doris.nereids.stats; + +import org.apache.doris.catalog.DatabaseIf; +import org.apache.doris.catalog.Env; +import org.apache.doris.catalog.OlapTable; +import org.apache.doris.catalog.TableIf; +import org.apache.doris.common.Config; +import org.apache.doris.datasource.CatalogIf; +import org.apache.doris.nereids.trees.plans.physical.PhysicalOlapScan; +import org.apache.doris.rpc.RpcException; +import org.apache.doris.statistics.hbo.PlanStatistics; +import org.apache.doris.statistics.hbo.ScanPlanStatistics; + +import org.apache.logging.log4j.LogManager; +import org.apache.logging.log4j.Logger; + +import java.util.ArrayList; +import java.util.Arrays; +import java.util.LinkedHashMap; +import java.util.LinkedHashSet; +import java.util.List; +import java.util.Locale; +import java.util.Map; +import java.util.Set; + +/** + * Whether a hbo entry may still be applied: the recorded data state of an entry (the baseline of + * every scan of its struct info, see {@link HboScanDescriptor}) compared with the data state of the + * query which wants to use it. + * + * <p>A hbo fingerprint does not contain any data state any more, so an entry keeps matching while a + * table grows. The read side therefore has to decide whether the injected row count still describes + * the data it was measured on: + * <ul> + * <li>{@link #STATE_LIVE}: the data the entry was measured on is still the current one, the entry + * is applied;</li> + * <li>{@link #STATE_DRIFTED}: the data moved, but by at most + * {@code Config.hbo_row_count_change_ratio} of the recorded row count (10% by default): the + * entry is applied, which is the point of the tolerance - a table which merely keeps growing + * (e.g. a few new rows in a partition) does not invalidate a measured row count;</li> + * <li>{@link #STATE_STALE}: the data moved further than that: the entry is <b>not</b> applied, and + * the query falls back to the optimizer estimation;</li> + * <li>{@link #STATE_UNKNOWN}: nothing can be compared (the entry records no row count, or the + * current one cannot be read / was obtained a different way), so there is nothing to judge: the + * entry is applied as it always was.</li> + * </ul> + * The recorded row count is never scaled: an entry either applies with the value the user measured + * or does not apply at all ("soft strategy" - a user who needs the exact recorded state in any case + * should not inject hbo statistics for that query). Setting + * {@code hbo_row_count_change_ratio} to 0 or less switches the verdict back to the strict semantics + * hbo had before the baseline existed: the recorded visible version has to still be the current one. + * + * <p>The rows are compared <b>before</b> the version, and only against a number which was obtained + * the same way ({@link HboScanDescriptor.RowsKind}): the visible version is a property of the table, + * so two different partition selections of one table share it while reading completely different + * amounts of data, and a measured row count must not be compared with one that was derived from the + * table average. Whenever nothing can be compared - a hand written struct literal, an entry whose + * rows were never measured, a current row count which cannot be read, or two numbers of different + * kinds - the state is {@link #STATE_UNKNOWN} and the entry is applied: a transient failure to read + * the catalog must not drop an entry the user injected. The version is only decisive in strict mode + * ({@code hbo_row_count_change_ratio <= 0}). + * + * <p>A scan is judged by the rows of its <b>selected</b> partitions, so pruning decides which data + * the entry depends on: an unrelated partition which grows does not invalidate a pruned entry. + * {@code HBO SHOW STATISTICS} and {@code HBO DELETE STALE STATISTICS} cannot run the query of an + * entry, so there the recorded state is compared with the table as a whole; an entry whose struct + * prunes partitions can therefore only be judged by rows when it was recorded as a whole table scan + * (its state is reported as unknown otherwise, and such an entry is kept unless the user asks for + * {@code OLDER_THAN}). + * + * <p>Learned (profile collected) entries carry the input table statistics of the run they were + * measured in, so they are judged by the same tolerance over those recorded rows; an entry injected + * by {@code HBO SET LEARNED STATISTICS} carries no such statistics and is applied as it always was. + */ +public class HboStructFreshness { + /** The recorded data state is still the current one. */ + public static final String STATE_LIVE = "live"; + /** The data moved, but within the tolerance of {@code Config.hbo_row_count_change_ratio}. */ + public static final String STATE_DRIFTED = "drifted"; + /** The data moved further than the tolerance: the entry must not be applied. */ + public static final String STATE_STALE = "stale"; + /** The entry records no data state (or it cannot be read any more): nothing to judge. */ + public static final String STATE_UNKNOWN = "unknown"; + + private static final Logger LOG = LogManager.getLogger(HboStructFreshness.class); + + private final String state; + // the recorded baseline and the current state, both rendered for display + private final String recorded; + private final String live; + // the row counts of the scan which decided the state, {@link HboScanDescriptor#UNKNOWN} when the + // state was decided by something else (e.g. a version which is not readable) + private final long recordedRows; + private final long liveRows; + + private HboStructFreshness(String state, String recorded, String live, long recordedRows, long liveRows) { + this.state = state; + this.recorded = recorded; + this.live = live; + this.recordedRows = recordedRows; + this.liveRows = liveRows; + } + + public String getState() { + return state; + } + + public boolean isStale() { + return STATE_STALE.equals(state); + } + + public boolean isUnknown() { + return STATE_UNKNOWN.equals(state); + } + + public boolean isDrifted() { + return STATE_DRIFTED.equals(state); + } + + /** The data state the entry was recorded in, e.g. {@code hbo_test.t1:v3,r1000,p1/5}. */ + public String getRecorded() { + return recorded; + } + + /** The data state the entry is compared with (empty when it cannot be read). */ + public String getLive() { + return live; + } + + /** + * The verdict as one token, printed after {@code used=} / {@code skipped=} in the explain + * annotation, e.g. {@code drifted(rows=1030,rec=1000,+3.0%)}. + */ + public String getSummary() { + if (!STATE_DRIFTED.equals(state) && !STATE_STALE.equals(state)) { + return state; + } + if (recordedRows == HboScanDescriptor.UNKNOWN || liveRows == HboScanDescriptor.UNKNOWN) { + return state; + } + return state + "(rows=" + liveRows + ",rec=" + recordedRows + "," + + changeText(recordedRows, liveRows) + ")"; + } + + /** The {@code now} part of {@code HBO SHOW STATISTICS}: the state read from the catalog. */ + public String getLiveDetail() { + if (live.isEmpty()) { + return "-"; + } + if (recordedRows == HboScanDescriptor.UNKNOWN || liveRows == HboScanDescriptor.UNKNOWN) { + return live; + } + return live + "," + changeText(recordedRows, liveRows); + } + + /** + * Whether the entry records any data state at all. An entry which does not (a struct literal + * typed by hand, or one copied from a plan whose tables were never reported) is always applied, + * so a caller does not have to compute the state of the current query to find that out. + */ + public static boolean hasRecordedDataState(String recordedCanonical) { + for (HboScanDescriptor scan : HboScanDescriptor.parseAll(recordedCanonical)) { + if (scan.hasVisibleVersion() || scan.hasScanRows()) { + return true; + } + } + return false; + } + + /** + * Judge an entry against the current query: the data state recorded in {@code recordedCanonical} + * (the struct info of the entry) against the state of {@code liveCanonical} (the struct info the + * read side just computed for the group). This is the decision of the read side. + */ + public static HboStructFreshness between(String recordedCanonical, GroupStructInfo liveStructInfo) { + return between(recordedCanonical, liveStructInfo.getScans()); + } + + /** The same comparison for a caller which has the scans of the current query only. */ + static HboStructFreshness between(String recordedCanonical, List<HboScanDescriptor> liveScans) { + List<HboScanDescriptor> recordedScans = HboScanDescriptor.parseAll(recordedCanonical); + String liveText = renderScans(liveScans); + if (recordedScans.isEmpty()) { + // an entry without a baseline (a hand written struct literal): nothing to judge + return new HboStructFreshness(STATE_UNKNOWN, "", liveText, + HboScanDescriptor.UNKNOWN, HboScanDescriptor.UNKNOWN); + } + if (recordedScans.size() != liveScans.size()) { + // the entry describes a different number of scans than the group: cannot be compared + return new HboStructFreshness(STATE_UNKNOWN, renderScans(recordedScans), liveText, + HboScanDescriptor.UNKNOWN, HboScanDescriptor.UNKNOWN); + } + String state = STATE_LIVE; + String recordedText = renderScans(recordedScans); + + long recordedRows = HboScanDescriptor.UNKNOWN; + long liveRows = HboScanDescriptor.UNKNOWN; + for (int i = 0; i < recordedScans.size(); i++) { + HboScanDescriptor recorded = recordedScans.get(i); + HboScanDescriptor live = liveScans.get(i); + String scanState; + if (!recorded.getTable().equals(live.getTable())) { + // the same fingerprint describes the same scans in the same order, so this cannot + // happen for an entry of this group: refuse to judge rather than guess + scanState = STATE_UNKNOWN; + } else { + scanState = stateOf(recorded, live); + } + if (worse(scanState, state).equals(scanState)) { + state = scanState; + recordedRows = recorded.getScanRows(); + liveRows = live.getScanRows(); + } + } + return new HboStructFreshness(state, recordedText, liveText, recordedRows, liveRows); + } + + /** + * Judge an entry against the catalog as a whole (used by {@code HBO SHOW STATISTICS} and + * {@code HBO DELETE STALE STATISTICS}, which cannot run the query of an entry): the recorded + * visible version and row count of every table against the current ones. + * + * <p>Row counts are only comparable when the scan read the whole table (no partition was pruned + * away), because only then the recorded rows are the rows of the table. + */ + public static HboStructFreshness of(String canonicalStructInfo) { + List<HboScanDescriptor> recordedScans = HboScanDescriptor.parseAll(canonicalStructInfo); + if (recordedScans.isEmpty()) { + return new HboStructFreshness(STATE_UNKNOWN, "", "", HboScanDescriptor.UNKNOWN, + HboScanDescriptor.UNKNOWN); + } + String state = STATE_LIVE; + // the current state of every table, de duplicated like the recorded side (a self join reads + // the same table twice, which would only repeat the same numbers) + Set<String> liveTexts = new LinkedHashSet<>(); + long recordedRows = HboScanDescriptor.UNKNOWN; + long liveRows = HboScanDescriptor.UNKNOWN; + for (HboScanDescriptor recorded : recordedScans) { + OlapTable table = resolveTable(recorded.getTable()); + if (table != null && tableReplaced(recorded, table)) { + // the name now resolves to another table (the recorded one was dropped and created again): + // the entry describes data which is gone, and no plan of the new table can produce its + // key, so judging that key by the replacement's state would keep a dead row as LIVE + liveTexts.add(recorded.getTable() + ":replaced#" + table.getId()); + if (worse(STATE_STALE, state).equals(STATE_STALE)) { + state = STATE_STALE; + recordedRows = HboScanDescriptor.UNKNOWN; + liveRows = HboScanDescriptor.UNKNOWN; + } + continue; + } + if (table == null) { + if (worse(STATE_UNKNOWN, state).equals(STATE_UNKNOWN)) { + state = STATE_UNKNOWN; + } + liveTexts.add(recorded.getTable() + ":-"); + continue; + } + if (table.getIndexIdList().size() > 1 && !recorded.isPartitionSelectionComplete()) { + // the table has more than one index (a rollup) and the entry records only a partition + // count, so this statement cannot tell which index its row count came from: judging it + // against the base index could be wrong in both directions, so it is left to a query + liveTexts.add(recorded.getTable() + ":-"); + if (worse(STATE_UNKNOWN, state).equals(STATE_UNKNOWN)) { + state = STATE_UNKNOWN; + } + continue; + } + long currentVersion = visibleVersionOf(table); + // the rows of the whole table, computed the same way the recorded rows of a whole table + // scan were computed, so the two numbers describe the same thing - and read from the index + // the entry read: a whole table query which the optimizer bound to an aggregate rollup + // recorded that rollup's rows, and comparing them with the base index would report an + // unchanged entry as stale (DELETE STALE would remove an entry a query can still apply) + HboScanDescriptor.Rows liveRowsOfTable = HboScanDescriptor.rowsOf(table, + new ArrayList<>(table.getPartitionIds()), indexIdOf(recorded, table)); + long currentRows = liveRowsOfTable.getRows(); + HboScanDescriptor.RowsKind currentRowsKind = liveRowsOfTable.getKind(); + liveTexts.add(recorded.getTable() + ":v" + currentVersion + + (currentRows == HboScanDescriptor.UNKNOWN ? "" + : (currentRowsKind == HboScanDescriptor.RowsKind.MEASURED ? ",r" : ",e") + + currentRows)); + String scanState; + if (!recorded.hasVisibleVersion() && !recorded.hasScanRows()) { + // nothing was recorded for this scan: it applies as it always was, nothing to judge + scanState = STATE_UNKNOWN; + currentRows = HboScanDescriptor.UNKNOWN; + } else if (threshold() <= 0) { + scanState = strictStateOf(recorded.hasVisibleVersion(), recorded.getVisibleVersion(), + currentVersion); + currentRows = HboScanDescriptor.UNKNOWN; + } else if (recorded.hasVisibleVersion() && currentVersion == recorded.getVisibleVersion() + && (currentRows == HboScanDescriptor.UNKNOWN + || currentRowsKind != recorded.getRowsKind() + || !recorded.hasScanRows() + || !recorded.isPartitionSelectionComplete())) { + // the visible version is a property of the whole table, so an unchanged version means + // that nothing was loaded into any of its partitions: the recorded state is still the + // current one, whatever the entry was measured on. (The read side can not use this + // shortcut: it knows which partitions the entry was measured on and which ones the + // query reads, and those can differ under one unchanged version.) + // Only a recorded scan which covers the whole table is compared below: its rows + // describe the same thing as the live rows of the table, and they can move under one + // unchanged version (an unreported partition uses TableStatsMeta, which ALTER COLUMN + // STATS ... ROW_COUNT updates), which the read side reports as stale as well. The + // rows of a pruned scan are not the rows of the table, so they are never compared. + scanState = STATE_LIVE; + } else if (!recorded.hasScanRows() || !recorded.isPartitionSelectionComplete()) { + // the entry records no row count at all (a struct literal of an older format, for + // example), or it records the rows of pruned partitions only, which are not the rows + // of the table: the state can only be judged by a query of that entry + scanState = STATE_UNKNOWN; + currentRows = HboScanDescriptor.UNKNOWN; + } else if (currentRows == HboScanDescriptor.UNKNOWN + || currentRowsKind != recorded.getRowsKind()) { + // the current row count cannot be read or was obtained differently than the recorded + // one, so the two are not comparable: the read side reports unknown for exactly this + // case and keeps applying the entry, and this scope must not disagree by calling it + // stale (DELETE STALE would then remove an entry which queries still use) + scanState = STATE_UNKNOWN; + currentRows = HboScanDescriptor.UNKNOWN; + } else if (recorded.getScanRows() == currentRows) { + scanState = STATE_LIVE; + } else { + scanState = changeRatio(recorded.getScanRows(), currentRows) <= threshold() + ? STATE_DRIFTED : STATE_STALE; + } + if (worse(scanState, state).equals(scanState) && scanState != STATE_LIVE) { + state = scanState; + recordedRows = scanState == STATE_UNKNOWN + ? HboScanDescriptor.UNKNOWN : recorded.getScanRows(); + liveRows = scanState == STATE_UNKNOWN ? HboScanDescriptor.UNKNOWN : currentRows; + } + } + return new HboStructFreshness(state, renderScans(recordedScans), String.join("; ", liveTexts), + recordedRows, liveRows); + } + + /** + * Whether the table a recorded scan names was dropped and created again under the same name: the token + * carries the id of the table it read, so a different id means the entry describes data which no + * longer exists. An entry of the older format carries none and cannot be judged. + */ + static boolean tableReplaced(HboScanDescriptor recorded, OlapTable table) { + return recorded.getTableId() != HboScanDescriptor.UNKNOWN + && recorded.getTableId() != table.getId(); + } + + /** + * The index a recorded scan of a whole table read: the one its token names, or the base index of the + * table when the token names none. A scan which reads a rollup records that index in its token, so + * the rows of the entry are compared with the rows of the index it was measured on. An index id + * which cannot be read (a struct info a user pasted with a hand written marker) leaves the catalog + * verdict unknown instead of failing the statement. + */ + static long indexIdOf(HboScanDescriptor recorded, OlapTable table) { + if (recorded.getIndexIdentity().isEmpty()) { + return table.getBaseIndexId(); + } + try { + return Long.parseLong(recorded.getIndexIdentity()); + } catch (NumberFormatException e) { + return HboScanDescriptor.UNKNOWN; + } + } + + /** + * Judge a learned (profile collected) entry: the rows its input tables had when it was measured + * against the rows of the scans of the current query. The publish path records one scan statistics + * per scan which feeds the published node, so this holds for an entry of any kind of node (see + * {@code HboStatsCalculator.applyLearnedStats}). The input tables are aggregated by table name, because + * a learned entry describes the scans of the node it was published for in plan order while the + * struct info orders them canonically. + * + * <p>This is a one way guard: it only rejects an entry when a comparable pair of row counts + * proves that the data moved beyond the tolerance. Everything which cannot be compared (an + * injected entry without input statistics, a different number of tables, a row count of a + * different kind) leaves the entry applicable, so a learned lookup which worked before this + * check keeps working. + */ + public static HboStructFreshness ofLearnedEntry(List<PlanStatistics> recordedInputStatistics, + GroupStructInfo liveStructInfo) { + return ofLearnedEntry(recordedInputStatistics, liveStructInfo.getScans()); + } + + /** The same guard for a caller which has the scans of the current query only. */ + static HboStructFreshness ofLearnedEntry(List<PlanStatistics> recordedInputStatistics, + List<HboScanDescriptor> liveScans) { + Map<String, Long> recordedRowsByTable = new LinkedHashMap<>(); + if (recordedInputStatistics != null) { + for (PlanStatistics input : recordedInputStatistics) { + if (!(input instanceof ScanPlanStatistics)) { + continue; + } + PhysicalOlapScan scan = ((ScanPlanStatistics) input).getScan(); + if (scan == null || scan.getTable() == null) { + continue; + } + String table = HboScanDescriptor.qualifiedNameOf(scan.getTable()); + // the input rows of a scan node are the rows it read, i.e. the number a catalog row + // count (which also ignores the predicates) can be compared with + recordedRowsByTable.merge(table, input.getInputRows(), Long::sum); Review Comment: learned HOB 是老版本的功能, 且老版本里已经是不可用状态. 不在这个pr里处理 -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
