Gabriel39 commented on code in PR #70708: URL: https://github.com/apache/doris/pull/70708#discussion_r4642211667
########## fe/fe-connector/fe-connector-mongodb/src/main/java/org/apache/doris/connector/mongodb/MongoRangePlanner.java: ########## @@ -0,0 +1,389 @@ +// 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.connector.mongodb; + +import org.apache.doris.connector.spi.ConnectorSession; +import org.apache.doris.connector.spi.ConnectorStatementScope; +import org.apache.doris.connector.spi.DorisConnectorException; +import org.apache.doris.connector.spi.scan.ConnectorScanContext; +import org.apache.doris.thrift.TMongoSnapshot; + +import org.bson.BsonValue; + +import java.util.HashMap; +import java.util.List; +import java.util.Map; +import java.util.Optional; +import java.util.Set; +import java.util.UUID; +import java.util.concurrent.CancellationException; +import java.util.concurrent.CompletableFuture; +import java.util.concurrent.ExecutionException; +import java.util.concurrent.Semaphore; +import java.util.concurrent.TimeUnit; +import java.util.concurrent.TimeoutException; +import java.util.function.Supplier; + +/** Complete planning precedes publication. Statement memos retain either that plan or its failure. */ +final class MongoRangePlanner { + static final int SAMPLE_ROWS = 4096; + static final int PROBE_TIMEOUT_MS = 2000; + private static final long MAX_PLAN_DESCRIPTOR_BYTES = 8L * 1024 * 1024; + private final String memoKey = "mongodb.split.plan:" + UUID.randomUUID(); + private final Supplier<MongoCatalogClient> client; + private final long probeTimeoutMs; + + MongoRangePlanner(Supplier<MongoCatalogClient> client, long probeTimeoutMs) { + this.client = client; + this.probeTimeoutMs = probeTimeoutMs; + } + + MongoSplitPlan plan(ConnectorSession session, MongoTableHandle table, TMongoSnapshot snapshot, + String preference, ConnectorScanContext context, long descriptorBytes) { + checkCancelled(session, context); + var options = session.getSessionProperties(); + String strategy = options.getOrDefault("mongodb_split_strategy", "single"); + if (!Set.of("single", "auto", "range").contains(strategy)) { + throw new IllegalArgumentException("Invalid mongodb_split_strategy"); + } + if ("single".equals(strategy)) { + return MongoSplitPlan.single(1, 1, "single_strategy"); + } + String field = options.getOrDefault("mongodb_split_field", ""); + boolean custom = !field.isEmpty(); + if (!custom) { + field = "_id"; + } + if (!MongoPredicateTranslator.safePath(field) || field.isEmpty() || field.length() > 256) { + throw new IllegalArgumentException("MongoDB split field must be an unambiguous top-level field"); + } + String selectedField = field; + Budget budget = session.getStatementScope().computeIfAbsent("mongodb.split.budget", Budget::new); + String tableKey = tableKey(session, table); + // Count fallback scans too: a view must not make a global custom field look unambiguous. + if (custom) { + budget.customField(tableKey); + } + Memo memo = session.getStatementScope().computeIfAbsent(tableKey, Memo::new); + return memo.get(session, context, () -> load(session, table, snapshot, preference, context, + strategy, selectedField, descriptorBytes, budget)); + } + + Optional<MongoSplitPlan> planned(ConnectorSession session, MongoTableHandle table) { + Memo memo = session.getStatementScope().getIfPresent(tableKey(session, table)); + return memo == null ? Optional.empty() : memo.completed(); + } + + void freezeDiagnostics(ConnectorSession session, MongoTableHandle table, Map<String, String> properties) { + session.getStatementScope().computeIfAbsent(tableKey(session, table) + ":diagnostics", + () -> Map.copyOf(properties)); + } + + Map<String, String> diagnostics(ConnectorSession session, MongoTableHandle table) { + Map<String, String> properties = session.getStatementScope() + .getIfPresent(tableKey(session, table) + ":diagnostics"); + return properties == null ? Map.of() : properties; + } + + private String tableKey(ConnectorSession session, MongoTableHandle table) { + // One provider can see several namespaces; never reuse another collection's bounds/admission. + return memoKey + ":" + session.getCatalogId() + ":" + table.getSchema().getVersion(); + } + + private MongoSplitPlan load(ConnectorSession session, MongoTableHandle table, TMongoSnapshot snapshot, + String preference, ConnectorScanContext context, String strategy, String field, + long descriptorBytes, Budget budget) { + int maximum = splitLimit(session); + validateSplitSizes(session); + int requested = Math.min(8, maximum); + // The default single strategy exits before this path; opted-in plans always reserve bytes. + assert descriptorBytes > 0 : "MongoDB range descriptor bytes must be positive"; + int descriptorMaximum = (int) Math.min(maximum, MAX_PLAN_DESCRIPTOR_BYTES / descriptorBytes); + if (descriptorMaximum == 0) { + throw new DorisConnectorException("MongoDB split descriptor exceeds planning byte budget"); + } + String source = snapshot == null ? "catalog:" + session.getCatalogId() + : snapshot.getSourceKind() + ":" + snapshot.getSourceId(); + if (!"collection".equals(table.getObjectKind())) { + return fallback(strategy, requested, "unsupported_object", source, descriptorBytes, budget); + } + if (snapshot == null || session.getStatementScope() == ConnectorStatementScope.NONE) { + return fallback(strategy, requested, "snapshot_required", source, descriptorBytes, budget); + } + if (descriptorMaximum == 1) { + return single(requested, maximum == 1 ? "split_limit" : "descriptor_budget", + source, descriptorBytes, budget); + } + long deadline = System.nanoTime() + TimeUnit.MILLISECONDS.toNanos( + Math.min(probeTimeoutMs, MongoSnapshot.remainingMillis(session))); + MongoSplitProbe probe; + try { + budget.acquireProbe(session, context, deadline); + try { + checkCancelled(session, context); + probe = client.get().probeSplits(table, field, preference, snapshot, SAMPLE_ROWS, + remaining(session, context, deadline), () -> checkCancelled(session, context)); + budget.sampled(probe.sampleBytes()); + } finally { + budget.probes.release(); + } + } catch (PlanningLimit limit) { + checkCancelled(session, context); + MongoSnapshot.remainingMillis(session); + // Only optional pre-publication tuning can fall back. Cancellation, authorization, + // history and source errors must never be converted into a successful full scan. + return fallback(strategy, requested, "probe_budget", source, descriptorBytes, budget); + } + checkCancelled(session, context); + MongoSnapshot.remainingMillis(session); + long targetBytes = targetBytes(session, probe.logicalBytes()); + long estimate = probe.logicalBytes() < 0 ? 8 + : Math.max(1, probe.logicalBytes() / targetBytes + (probe.logicalBytes() % targetBytes == 0 ? 0 : 1)); + requested = (int) Math.min(maximum, estimate); + if (!probe.indexed()) { + return fallback(strategy, requested, "missing_index", source, descriptorBytes, budget); + } + String family = family(probe.samples()); + if (family.isEmpty()) { + return probe.samples().isEmpty() + ? single(requested, "empty_sample", source, descriptorBytes, budget) + : fallback(strategy, requested, "unsupported_type", source, descriptorBytes, budget); + } + if ("auto".equals(strategy) && probe.logicalBytes() < 0) { + return single(requested, "unknown_complement_cost", source, descriptorBytes, budget); + } + // Division avoids overflow when a valid shared split-size setting is near Long.MAX_VALUE. + if ("auto".equals(strategy) && probe.logicalBytes() / targetBytes < 2) { + return single(requested, "small_collection", source, descriptorBytes, budget); + } + int admitted = budget.admit(source, Math.min(requested, descriptorMaximum), descriptorBytes); + try { + var values = probe.samples().stream().filter(v -> family.equals(family(v))).toList(); + MongoSplitPlan plan = MongoSplitPlan.ranges(field, family, values, requested, admitted); + budget.refund(source, admitted - plan.size(), descriptorBytes); + return plan; + } catch (RuntimeException failure) { + budget.refund(source, admitted, descriptorBytes); + throw failure; + } + } + + static int splitLimit(ConnectorSession session) { + long limit = number(session.getSessionProperties(), "max_file_split_num", 100000); + // Mongo freezes all boundaries even when transport is streaming; that retained plan must + // honor the shared count cap too. Nonpositive means no user cap, as on other file scans. + return limit <= 0 ? 256 : (int) Math.min(256, limit); + } + + private static void validateSplitSizes(ConnectorSession session) { + targetBytes(session, -1); + } + + static long targetBytes(ConnectorSession session, long logicalBytes) { + var options = session.getSessionProperties(); + long specified = number(options, "file_split_size", 0); + if (specified > 0) { + return specified; + } + long regular = number(options, "max_file_split_size", 64L * 1024 * 1024); + long initial = number(options, "max_initial_file_split_size", 32L * 1024 * 1024); + long initialCount = number(options, "max_initial_file_split_num", 200); + if (regular <= 0 || initial <= 0) { + throw new IllegalArgumentException("MongoDB split sizes must be positive"); + } + // Match the shared scan-size heuristic without overflowing regular * initialCount. + return logicalBytes < 0 || initialCount <= 0 || logicalBytes / regular >= initialCount + ? regular : initial; + } + + static void checkCancelled(ConnectorSession session, ConnectorScanContext context) { + context.checkCancelled(); + // Synchronous final planning may have no registered scan context. Observe the same query + // task-group lifetime as asynchronous split producers before every remote cursor advance. + MongoSnapshot.checkQueryCancelled(session); + } + + static String family(BsonValue value) { + return value.isInt32() || value.isInt64() ? "integer" + : value.isObjectId() ? "objectId" : value.isDateTime() ? "date" : ""; + } + + static String family(List<BsonValue> samples) { + Map<String, Integer> counts = new HashMap<>(); + for (BsonValue value : samples) { + String family = family(value); + if (!family.isEmpty()) { + counts.merge(family, 1, Integer::sum); + } + } + // Deterministic ties affect balance only; the complete partitions still cover every source type. + return counts.keySet().stream().sorted().max(java.util.Comparator.comparingInt(counts::get)).orElse(""); + } + + private static MongoSplitPlan fallback(String strategy, int requested, String reason, + String source, long descriptorBytes, Budget budget) { + if ("range".equals(strategy)) { + throw new IllegalArgumentException("MongoDB range planning unavailable: " + reason); + } + return single(requested, reason, source, descriptorBytes, budget); + } + + private static MongoSplitPlan single(int requested, String reason, String source, + long descriptorBytes, Budget budget) { + // A successful fallback still publishes a descriptor and a source scan. For example, + // 257 empty collection scans must not bypass the same 256-partition pinned-source cap. + budget.admit(source, 1, descriptorBytes); + return MongoSplitPlan.single(requested, 1, reason); + } + + static long remaining(ConnectorSession session, ConnectorScanContext context, long deadline) { + checkCancelled(session, context); + long query = MongoSnapshot.remainingMillis(session); + long left = deadline - System.nanoTime(); + if (left <= 0) { + throw new PlanningLimit(); + } + return Math.min(query, Math.max(1, TimeUnit.NANOSECONDS.toMillis(left))); + } + + private static long number(Map<String, String> options, String name, long fallback) { + try { + return Long.parseLong(options.getOrDefault(name, Long.toString(fallback))); + } catch (NumberFormatException malformed) { + throw new IllegalArgumentException("Invalid split setting: " + name); + } + } + + static final class PlanningLimit extends RuntimeException { + PlanningLimit() { + super("MongoDB split probe budget exceeded"); + } + } + + static final class Budget { + private final Semaphore probes = new Semaphore(2); + private final Map<String, Integer> sourceSplits = new HashMap<>(); + private int splits; + private long bytes; + private long descriptorBytes; + private final Map<String, Long> sourceDescriptorBytes = new HashMap<>(); + private String fieldOwner; + + synchronized void customField(String owner) { + if (fieldOwner != null && !fieldOwner.equals(owner)) { + throw new IllegalArgumentException("mongodb_split_field requires exactly one MongoDB scan"); + } + fieldOwner = owner; + } + + synchronized int admit(String source, int requested, long perRangeBytes) { + long remainingBytes = Math.min(64L * 1024 * 1024 - descriptorBytes, + 32L * 1024 * 1024 - sourceDescriptorBytes.getOrDefault(source, 0L)); + int remaining = (int) Math.min(Math.min(4096 - splits, 256 - sourceSplits.getOrDefault(source, 0)), + remainingBytes / perRangeBytes); + if (remaining < 1) { + throw new DorisConnectorException("MongoDB statement/source split budget exhausted"); + } + int count = Math.min(requested, remaining); + splits += count; + descriptorBytes += count * perRangeBytes; + sourceSplits.merge(source, count, Integer::sum); + sourceDescriptorBytes.merge(source, count * perRangeBytes, Long::sum); + return count; + } + + synchronized void refund(String source, int count, long perRangeBytes) { + splits -= count; + descriptorBytes -= count * perRangeBytes; + sourceSplits.merge(source, -count, Integer::sum); + sourceDescriptorBytes.merge(source, -count * perRangeBytes, Long::sum); + } + + synchronized void sampled(long count) { + if (count < 0 || count > 1024 * 1024 || bytes + count > 16L * 1024 * 1024) { + throw new PlanningLimit(); + } + bytes += count; + } + + void acquireProbe(ConnectorSession session, ConnectorScanContext context, long deadline) { + try { + while (!probes.tryAcquire(Math.min(50, remaining(session, context, deadline)), TimeUnit.MILLISECONDS)) { + checkCancelled(session, context); + } + } catch (InterruptedException interrupted) { + Thread.currentThread().interrupt(); + throw new CancellationException("MongoDB split planning interrupted"); + } + } + } + + private static final class Memo { + private CompletableFuture<MongoSplitPlan> value; + + synchronized Optional<MongoSplitPlan> completed() { + return value != null && value.isDone() && !value.isCompletedExceptionally() + ? Optional.of(value.getNow(null)) : Optional.empty(); + } + + MongoSplitPlan get(ConnectorSession session, ConnectorScanContext context, Supplier<MongoSplitPlan> loader) { + CompletableFuture<MongoSplitPlan> selected; + boolean owner; + synchronized (this) { + owner = value == null; + if (owner) { + value = new CompletableFuture<>(); + } + selected = value; + } + if (owner) { + try { + selected.complete(loader.get()); + } catch (Throwable failure) { + // A fatal loader failure must wake every waiter instead of leaving the memo pending. + selected.completeExceptionally(failure); + if (failure instanceof Error fatal) { + throw fatal; + } + } + } + while (true) { + checkCancelled(session, context); + try { + // An offline auto fallback has no statement pin/deadline. Its private memo + // was completed by this caller; only live statement waiters need timed polling. + return session.getStatementScope() == ConnectorStatementScope.NONE ? selected.get() + : selected.get(Math.min(50, MongoSnapshot.remainingMillis(session)), TimeUnit.MILLISECONDS); Review Comment: Fixed in `41a2f893579`. Completed futures are read before consulting the statement deadline; pending live waiters still use timed polling. Cancellation and thread-interruption checks remain before both paths. A deterministic test expires the actual statement state without sleeps and separately verifies the completed plan and original source exception identity. Both checks fail on the baseline and pass with the fix; the full 169-test MongoDB suite passes. ########## fe/fe-connector/fe-connector-mongodb/src/main/java/org/apache/doris/connector/mongodb/MongoScanPlanProvider.java: ########## @@ -132,28 +252,71 @@ public Map<String, String> getScanNodeProperties(ConnectorSession session, Conne Map<String, String> result = new LinkedHashMap<>(); var policies = session.getSessionProperties(); String concern = properties.resolveReadConcern(policies.getOrDefault("mongodb_read_concern", "inherit")); - String preference = properties.resolveReadPreference( - policies.getOrDefault("mongodb_read_preference", "inherit")); - var snapshot = MongoSnapshot.prepare(session, properties, table, client, concern, preference, - 1000L * bounded(policies, "query_timeout", 300, 1, Integer.MAX_VALUE / 1000)); + var pushed = predicate(session, table, filter); + var pruned = projection(session, table, columns); result.put("read_concern", concern); - result.put("snapshot", snapshot == null ? "none" : snapshot.getTimestampSeconds() - + ":" + snapshot.getTimestampIncrement()); result.put("object_kind", table.getObjectKind()); result.put("schema_source", table.getSchema().getSource()); result.put("schema_version", table.getSchema().getVersion()); - result.put("predicate_pushdown", predicate(session, table, filter).kind().name()); - result.put("projection_pushdown", projection(session, table, columns).isEmpty() ? "false" : "true"); - return result; + result.put("predicate_pushdown", pushed.kind().name()); + result.put("projection_pushdown", pruned.isEmpty() ? "false" : "true"); + // initSchemaParams reads properties BEFORE WHERE/projection are attached. Source probes + // and byte admission here would reject a normal large predicate and ignore KILL QUERY. + // Resolve diagnostics lazily from the final cancellable plan, without doing any source I/O. + return new AbstractMap<>() { Review Comment: Fixed in `41a2f893579`. get(Object) serves static and absent keys directly from the captured base map, while diagnostic keys retain the final lazy snapshot. A baseline-failing counter test reduces statement-scope reads for four such lookups from 12 to zero, then compares every final entry with get(key). Existing predicate/projection and one-generation EXPLAIN tests still pass, as do all 169 MongoDB tests and the complete split SQL suite. -- 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]
