@fluidframework/odsp-driver 1.4.0-121020 → 2.0.0-dev-rc.1.0.0.225277

package/src/odspDocumentStorageServiceBase.ts

@@ -3,266 +3,291 @@
  * Licensed under the MIT License.
  */
 
-import { assert } from "@fluidframework/common-utils";
-import { IDocumentStorageService, ISummaryContext, LoaderCachingPolicy } from "@fluidframework/driver-definitions";
+import { assert } from "@fluidframework/core-utils";
+import {
+	IDocumentStorageService,
+	IDocumentStorageServicePolicies,
+	ISummaryContext,
+	LoaderCachingPolicy,
+	FiveDaysMs,
+	FetchSource,
+} from "@fluidframework/driver-definitions";
 import * as api from "@fluidframework/protocol-definitions";
-import { defaultCacheExpiryTimeoutMs } from "./epochTracker";
+import { IConfigProvider } from "@fluidframework/telemetry-utils";
 import { ISnapshotContents } from "./odspPublicUtils";
 
-/* eslint-disable max-len */
+const maximumCacheDurationMs: FiveDaysMs = 432000000; // 5 * 24 * 60 * 60 * 1000 = 5 days in ms
 
 class BlobCache {
-    // Save the timeout so we can cancel and reschedule it as needed
-    private blobCacheTimeout: ReturnType<typeof setTimeout> | undefined;
-    // If the defer flag is set when the timeout fires, we'll reschedule rather than clear immediately
-    // This deferral approach is used (rather than clearing/resetting the timer) as current calling patterns trigger
-    // too many calls to setTimeout/clearTimeout.
-    private deferBlobCacheClear: boolean = false;
-
-    private readonly _blobCache: Map<string, ArrayBuffer> = new Map();
-
-    // Tracks all blob IDs evicted from cache
-    private readonly blobsEvicted: Set<string> = new Set();
-
-    // Initial time-out to purge data from cache
-    // If this time out is very small, then we purge blobs from cache too soon and that results in a lot of
-    // requests to storage, which brings down perf and may trip protection limits causing 429s
-    private blobCacheTimeoutDuration = 2 * 60 * 1000;
-
-    // SPO does not keep old snapshots around for long, so we are running chances of not
-    // being able to rehydrate data store / DDS in the future if we purge anything (and with blob de-duping,
-    // even if blob read by runtime, it could be read again in the future)
-    // So for now, purging is disabled.
-    private readonly purgeEnabled = false;
-
-    public get value() {
-        return this._blobCache;
-    }
-
-    public addBlobs(blobs: Map<string, ArrayBuffer>) {
-        blobs.forEach((value, blobId) => {
-            this._blobCache.set(blobId, value);
-        });
-        // Reset the timer on cache set
-        this.scheduleClearBlobsCache();
-    }
-
-    /**
-     * Schedule a timer for clearing the blob cache or defer the current one.
-     */
-    private scheduleClearBlobsCache() {
-        if (this.blobCacheTimeout !== undefined) {
-            // If we already have an outstanding timer, just signal that we should defer the clear
-            this.deferBlobCacheClear = true;
-        } else if (this.purgeEnabled) {
-            // If we don't have an outstanding timer, set a timer
-            // When the timer runs out, we'll decide whether to proceed with the cache clear or reset the timer
-            const clearCacheOrDefer = () => {
-                this.blobCacheTimeout = undefined;
-                if (this.deferBlobCacheClear) {
-                    this.deferBlobCacheClear = false;
-                    this.scheduleClearBlobsCache();
-                } else {
-                    // NOTE: Slightly better algorithm here would be to purge either only big blobs,
-                    // or sort them by size and purge enough big blobs to leave only 256Kb of small blobs in cache
-                    // Purging is optimizing memory footprint. But count controls potential number of storage requests
-                    // We want to optimize both - memory footprint and number of future requests to storage.
-                    // Note that Container can realize data store or DDS on-demand at any point in time, so we do not
-                    // control when blobs will be used.
-                    this._blobCache.forEach((_, blobId) => this.blobsEvicted.add(blobId));
-                    this._blobCache.clear();
-                }
-            };
-            this.blobCacheTimeout = setTimeout(clearCacheOrDefer, this.blobCacheTimeoutDuration);
-            // any future storage reads that get into the cache should be cleared from cache rather quickly -
-            // there is not much value in keeping them longer
-            this.blobCacheTimeoutDuration = 10 * 1000;
-        }
-    }
-
-    public getBlob(blobId: string) {
-        // Reset the timer on attempted cache read
-        this.scheduleClearBlobsCache();
-        const blobContent = this._blobCache.get(blobId);
-        const evicted = this.blobsEvicted.has(blobId);
-        return { blobContent, evicted };
-    }
-
-    public setBlob(blobId: string, blob: ArrayBuffer) {
-        // This API is called as result of cache miss and reading blob from storage.
-        // Runtime never reads same blob twice.
-        // The only reason we may get read request for same blob is blob de-duping in summaries.
-        // Note that the bigger the size, the less likely blobs are the same, so there is very little benefit of caching big blobs.
-        // Images are the only exception - user may insert same image twice. But we currently do not de-dup them - only snapshot
-        // blobs are de-duped.
-        const size = blob.byteLength;
-        if (size < 256 * 1024) {
-            // Reset the timer on cache set
-            this.scheduleClearBlobsCache();
-            return this._blobCache.set(blobId, blob);
-        } else {
-            // we evicted it here by not caching.
-            this.blobsEvicted.add(blobId);
-        }
-    }
+	// Save the timeout so we can cancel and reschedule it as needed
+	private blobCacheTimeout: ReturnType<typeof setTimeout> | undefined;
+	// If the defer flag is set when the timeout fires, we'll reschedule rather than clear immediately
+	// This deferral approach is used (rather than clearing/resetting the timer) as current calling patterns trigger
+	// too many calls to setTimeout/clearTimeout.
+	private deferBlobCacheClear: boolean = false;
+
+	private readonly _blobCache: Map<string, ArrayBuffer> = new Map();
+
+	// Tracks all blob IDs evicted from cache
+	private readonly blobsEvicted: Set<string> = new Set();
+
+	// Initial time-out to purge data from cache
+	// If this time out is very small, then we purge blobs from cache too soon and that results in a lot of
+	// requests to storage, which brings down perf and may trip protection limits causing 429s
+	private blobCacheTimeoutDuration = 2 * 60 * 1000;
+
+	// SPO does not keep old snapshots around for long, so we are running chances of not
+	// being able to rehydrate data store / DDS in the future if we purge anything (and with blob de-duping,
+	// even if blob read by runtime, it could be read again in the future)
+	// So for now, purging is disabled.
+	private readonly purgeEnabled = false;
+
+	public get value() {
+		return this._blobCache;
+	}
+
+	public addBlobs(blobs: Map<string, ArrayBuffer>) {
+		blobs.forEach((value, blobId) => {
+			this._blobCache.set(blobId, value);
+		});
+		// Reset the timer on cache set
+		this.scheduleClearBlobsCache();
+	}
+
+	/**
+	 * Schedule a timer for clearing the blob cache or defer the current one.
+	 */
+	private scheduleClearBlobsCache() {
+		if (this.blobCacheTimeout !== undefined) {
+			// If we already have an outstanding timer, just signal that we should defer the clear
+			this.deferBlobCacheClear = true;
+		} else if (this.purgeEnabled) {
+			// If we don't have an outstanding timer, set a timer
+			// When the timer runs out, we'll decide whether to proceed with the cache clear or reset the timer
+			const clearCacheOrDefer = () => {
+				this.blobCacheTimeout = undefined;
+				if (this.deferBlobCacheClear) {
+					this.deferBlobCacheClear = false;
+					this.scheduleClearBlobsCache();
+				} else {
+					// NOTE: Slightly better algorithm here would be to purge either only big blobs,
+					// or sort them by size and purge enough big blobs to leave only 256Kb of small blobs in cache
+					// Purging is optimizing memory footprint. But count controls potential number of storage requests
+					// We want to optimize both - memory footprint and number of future requests to storage.
+					// Note that Container can realize data store or DDS on-demand at any point in time, so we do not
+					// control when blobs will be used.
+					this._blobCache.forEach((_, blobId) => this.blobsEvicted.add(blobId));
+					this._blobCache.clear();
+				}
+			};
+			this.blobCacheTimeout = setTimeout(clearCacheOrDefer, this.blobCacheTimeoutDuration);
+			// any future storage reads that get into the cache should be cleared from cache rather quickly -
+			// there is not much value in keeping them longer
+			this.blobCacheTimeoutDuration = 10 * 1000;
+		}
+	}
+
+	public getBlob(blobId: string) {
+		// Reset the timer on attempted cache read
+		this.scheduleClearBlobsCache();
+		const blobContent = this._blobCache.get(blobId);
+		const evicted = this.blobsEvicted.has(blobId);
+		return { blobContent, evicted };
+	}
+
+	public setBlob(blobId: string, blob: ArrayBuffer) {
+		// This API is called as result of cache miss and reading blob from storage.
+		// Runtime never reads same blob twice.
+		// The only reason we may get read request for same blob is blob de-duping in summaries.
+		// Note that the bigger the size, the less likely blobs are the same, so there is very little benefit of caching big blobs.
+		// Images are the only exception - user may insert same image twice. But we currently do not de-dup them - only snapshot
+		// blobs are de-duped.
+		const size = blob.byteLength;
+		if (size < 256 * 1024) {
+			// Reset the timer on cache set
+			this.scheduleClearBlobsCache();
+			return this._blobCache.set(blobId, blob);
+		} else {
+			// we evicted it here by not caching.
+			this.blobsEvicted.add(blobId);
+		}
+	}
 }
 
 export abstract class OdspDocumentStorageServiceBase implements IDocumentStorageService {
-    readonly policies = {
-        // By default, ODSP tells the container not to prefetch/cache.
-        caching: LoaderCachingPolicy.NoCaching,
-
-        // ODSP storage works better if it has less number of blobs / edges
-        // Runtime creating many small blobs results in sub-optimal perf.
-        // 2K seems like the sweat spot:
-        // The smaller the number, less blobs we aggregate. Most storages are very likely to have notion
-        // of minimal "cluster" size, so having small blobs is wasteful
-        // At the same time increasing the limit ensure that more blobs with user content are aggregated,
-        // reducing possibility for de-duping of same blobs (i.e. .attributes rolled into aggregate blob
-        // are not reused across data stores, or even within data store, resulting in duplication of content)
-        // Note that duplication of content should not have significant impact for bytes over wire as
-        // compression of http payload mostly takes care of it, but it does impact storage size and in-memory sizes.
-        minBlobSize: 2048,
-        maximumCacheDurationMs: defaultCacheExpiryTimeoutMs,
-    };
-
-    protected readonly commitCache: Map<string, api.ISnapshotTree> = new Map();
-
-    private readonly attributesBlobHandles: Set<string> = new Set();
-
-    private _ops: api.ISequencedDocumentMessage[] | undefined;
-
-    private _snapshotSequenceNumber: number | undefined;
-
-    protected readonly blobCache = new BlobCache();
-
-    public set ops(ops: api.ISequencedDocumentMessage[] | undefined) {
-        assert(this._ops === undefined, 0x0a5 /* "Trying to set ops when they are already set!" */);
-        this._ops = ops;
-    }
-
-    public get ops(): api.ISequencedDocumentMessage[] | undefined {
-        return this._ops;
-    }
-
-    public get snapshotSequenceNumber() {
-        return this._snapshotSequenceNumber;
-    }
-
-    public get repositoryUrl(): string {
-        return "";
-    }
-
-    public abstract createBlob(file: ArrayBufferLike): Promise<api.ICreateBlobResponse>;
-
-    private async readBlobCore(blobId: string): Promise<ArrayBuffer> {
-        const { blobContent, evicted } = this.blobCache.getBlob(blobId);
-        return blobContent ?? this.fetchBlobFromStorage(blobId, evicted);
-    }
-
-    protected abstract fetchBlobFromStorage(blobId: string, evicted: boolean): Promise<ArrayBuffer>;
-
-    public async readBlob(blobId: string): Promise<ArrayBufferLike> {
-        return this.readBlobCore(blobId);
-    }
-
-    public async getSnapshotTree(version?: api.IVersion, scenarioName?: string): Promise<api.ISnapshotTree | null> {
-        let id: string;
-        if (!version || !version.id) {
-            const versions = await this.getVersions(null, 1, scenarioName);
-            if (!versions || versions.length === 0) {
-                return null;
-            }
-            id = versions[0].id;
-        } else {
-            id = version.id;
-        }
-
-        const snapshotTree = await this.readTree(id, scenarioName);
-        if (!snapshotTree) {
-            return null;
-        }
-
-        if (snapshotTree.blobs) {
-            const attributesBlob = snapshotTree.blobs.attributes;
-            if (attributesBlob) {
-                this.attributesBlobHandles.add(attributesBlob);
-            }
-        }
-
-        // When we upload the container snapshot, we upload appTree in ".app" and protocol tree in ".protocol"
-        // So when we request the snapshot we get ".app" as tree and not as commit node as in the case just above.
-        const appTree = snapshotTree.trees[".app"];
-        const protocolTree = snapshotTree.trees[".protocol"];
-
-        return this.combineProtocolAndAppSnapshotTree(appTree, protocolTree);
-    }
-
-    public abstract getVersions(blobid: string | null, count: number, scenarioName?: string): Promise<api.IVersion[]>;
-
-    public abstract uploadSummaryWithContext(summary: api.ISummaryTree, context: ISummaryContext): Promise<string>;
-
-    public async downloadSummary(commit: api.ISummaryHandle): Promise<api.ISummaryTree> {
-        throw new Error("Not implemented yet");
-    }
-
-    protected setRootTree(id: string, tree: api.ISnapshotTree) {
-        this.commitCache.set(id, tree);
-    }
-
-    protected initBlobsCache(blobs: Map<string, ArrayBuffer>) {
-        this.blobCache.addBlobs(blobs);
-    }
-
-    private async readTree(id: string, scenarioName?: string): Promise<api.ISnapshotTree | null> {
-        let tree = this.commitCache.get(id);
-        if (!tree) {
-            tree = await this.fetchTreeFromSnapshot(id, scenarioName);
-        }
-
-        return tree ?? null;
-    }
-
-    protected abstract fetchTreeFromSnapshot(id: string, scenarioName?: string): Promise<api.ISnapshotTree | undefined>;
-
-    private combineProtocolAndAppSnapshotTree(
-        hierarchicalAppTree: api.ISnapshotTree,
-        hierarchicalProtocolTree: api.ISnapshotTree,
-    ) {
-        const summarySnapshotTree: api.ISnapshotTree = {
-            blobs: {
-                ...hierarchicalAppTree.blobs,
-            },
-            trees: {
-                ...hierarchicalAppTree.trees,
-                // the app tree could have a .protocol
-                // in that case we want to server protocol to override it
-                ".protocol": hierarchicalProtocolTree,
-            },
-        };
-
-        return summarySnapshotTree;
-    }
-
-    protected initializeFromSnapshot(odspSnapshotCacheValue: ISnapshotContents): string | undefined {
-        this._snapshotSequenceNumber = odspSnapshotCacheValue.sequenceNumber;
-        const { snapshotTree, blobs, ops } = odspSnapshotCacheValue;
-
-        // id should be undefined in case of just ops in snapshot.
-        let id: string | undefined;
-        if (snapshotTree) {
-            id = snapshotTree.id;
-            assert(id !== undefined, 0x221 /* "Root tree should contain the id" */);
-            this.setRootTree(id, snapshotTree);
-        }
-
-        if (blobs) {
-            this.initBlobsCache(blobs);
-        }
-
-        this.ops = ops;
-        return id;
-    }
+	readonly policies: IDocumentStorageServicePolicies;
+
+	constructor(config: IConfigProvider) {
+		// We circumvent the restrictions on the policy only when using this TestOverride setting,
+		// which also applies to the code that reads from the cache in epochTracker.ts
+		// This may result in files created for testing being unusable in production sessions,
+		// due to the GC code guarding against this policy changing over the lifetime of a file.
+		const maximumCacheDurationMsInEffect = (
+			config.getBoolean("Fluid.Driver.Odsp.TestOverride.DisableSnapshotCache")
+				? 0
+				: maximumCacheDurationMs
+		) as FiveDaysMs;
+
+		this.policies = {
+			// By default, ODSP tells the container not to prefetch/cache.
+			caching: LoaderCachingPolicy.NoCaching,
+			maximumCacheDurationMs: maximumCacheDurationMsInEffect,
+		};
+	}
+	protected readonly commitCache: Map<string, api.ISnapshotTree> = new Map();
+
+	private readonly attributesBlobHandles: Set<string> = new Set();
+
+	private _ops: api.ISequencedDocumentMessage[] | undefined;
+
+	private _snapshotSequenceNumber: number | undefined;
+
+	protected readonly blobCache = new BlobCache();
+
+	public set ops(ops: api.ISequencedDocumentMessage[] | undefined) {
+		assert(this._ops === undefined, 0x0a5 /* "Trying to set ops when they are already set!" */);
+		this._ops = ops;
+	}
+
+	public get ops(): api.ISequencedDocumentMessage[] | undefined {
+		return this._ops;
+	}
+
+	public get snapshotSequenceNumber() {
+		return this._snapshotSequenceNumber;
+	}
+
+	public readonly repositoryUrl = "";
+
+	public abstract createBlob(file: ArrayBufferLike): Promise<api.ICreateBlobResponse>;
+
+	private async readBlobCore(blobId: string): Promise<ArrayBuffer> {
+		const { blobContent, evicted } = this.blobCache.getBlob(blobId);
+		return blobContent ?? this.fetchBlobFromStorage(blobId, evicted);
+	}
+
+	protected abstract fetchBlobFromStorage(blobId: string, evicted: boolean): Promise<ArrayBuffer>;
+
+	public async readBlob(blobId: string): Promise<ArrayBufferLike> {
+		return this.readBlobCore(blobId);
+	}
+
+	public async getSnapshotTree(
+		version?: api.IVersion,
+		scenarioName?: string,
+		// eslint-disable-next-line @rushstack/no-new-null
+	): Promise<api.ISnapshotTree | null> {
+		let id: string;
+		if (!version?.id) {
+			const versions = await this.getVersions(null, 1, scenarioName);
+			if (!versions || versions.length === 0) {
+				return null;
+			}
+			id = versions[0].id;
+		} else {
+			id = version.id;
+		}
+
+		const snapshotTree = await this.readTree(id, scenarioName);
+		if (!snapshotTree) {
+			return null;
+		}
+
+		if (snapshotTree.blobs) {
+			const attributesBlob = snapshotTree.blobs.attributes;
+			if (attributesBlob) {
+				this.attributesBlobHandles.add(attributesBlob);
+			}
+		}
+
+		// When we upload the container snapshot, we upload appTree in ".app" and protocol tree in ".protocol"
+		// So when we request the snapshot we get ".app" as tree and not as commit node as in the case just above.
+		const appTree = snapshotTree.trees[".app"];
+		const protocolTree = snapshotTree.trees[".protocol"];
+
+		return this.combineProtocolAndAppSnapshotTree(appTree, protocolTree);
+	}
+
+	public abstract getVersions(
+		// eslint-disable-next-line @rushstack/no-new-null
+		blobid: string | null,
+		count: number,
+		scenarioName?: string,
+		fetchSource?: FetchSource,
+	): Promise<api.IVersion[]>;
+
+	public abstract uploadSummaryWithContext(
+		summary: api.ISummaryTree,
+		context: ISummaryContext,
+	): Promise<string>;
+
+	public async downloadSummary(commit: api.ISummaryHandle): Promise<api.ISummaryTree> {
+		throw new Error("Not implemented yet");
+	}
+
+	protected setRootTree(id: string, tree: api.ISnapshotTree) {
+		this.commitCache.set(id, tree);
+	}
+
+	protected initBlobsCache(blobs: Map<string, ArrayBuffer>) {
+		this.blobCache.addBlobs(blobs);
+	}
+
+	private async readTree(id: string, scenarioName?: string): Promise<api.ISnapshotTree | null> {
+		let tree = this.commitCache.get(id);
+		if (!tree) {
+			tree = await this.fetchTreeFromSnapshot(id, scenarioName);
+		}
+
+		return tree ?? null;
+	}
+
+	protected abstract fetchTreeFromSnapshot(
+		id: string,
+		scenarioName?: string,
+	): Promise<api.ISnapshotTree | undefined>;
+
+	private combineProtocolAndAppSnapshotTree(
+		hierarchicalAppTree: api.ISnapshotTree,
+		hierarchicalProtocolTree: api.ISnapshotTree,
+	) {
+		const summarySnapshotTree: api.ISnapshotTree = {
+			blobs: {
+				...hierarchicalAppTree.blobs,
+			},
+			trees: {
+				...hierarchicalAppTree.trees,
+				// the app tree could have a .protocol
+				// in that case we want to server protocol to override it
+				".protocol": hierarchicalProtocolTree,
+			},
+		};
+
+		return summarySnapshotTree;
+	}
+
+	protected initializeFromSnapshot(
+		odspSnapshotCacheValue: ISnapshotContents,
+		cacheOps: boolean = true,
+	): string | undefined {
+		this._snapshotSequenceNumber = odspSnapshotCacheValue.sequenceNumber;
+		const { snapshotTree, blobs, ops } = odspSnapshotCacheValue;
+
+		// id should be undefined in case of just ops in snapshot.
+		let id: string | undefined;
+		if (snapshotTree) {
+			id = snapshotTree.id;
+			assert(id !== undefined, 0x221 /* "Root tree should contain the id" */);
+			this.setRootTree(id, snapshotTree);
+		}
+
+		if (blobs) {
+			this.initBlobsCache(blobs);
+		}
+
+		if (cacheOps) {
+			this.ops = ops;
+		}
+		return id;
+	}
 }
-
-/* eslint-enable max-len */