Add blob split and splice API (#282)

Depending on the software project, possibly large binary artifacts need to be
downloaded from or uploaded to the remote CAS. Examples are executables with
debug information, comprehensive libraries, or even whole file system images.
Such artifacts generate a lot of traffic when downloaded or uploaded.

The blob split API allows to split such artifacts into chunks at the remote
side, to fetch only those parts that are locally missing, and finally to
locally assemble the requested blob from its chunks.

The blob splice API allows to split such artifacts into chunks locally, to
upload only those parts that are remotely missing, and finally to remotely
splice the requested blob from its chunks.

Since only the binary differences from the last download/upload are
fetched/uploaded, the blob split and splice API can significantly save network
traffic between server and client.

Author: roloffs

build/bazel/remote/execution/v2/remote_execution.proto

@@ -430,6 +430,103 @@ service ContentAddressableStorage {
   rpc GetTree(GetTreeRequest) returns (stream GetTreeResponse) {
     option (google.api.http) = { get: "/v2/{instance_name=**}/blobs/{root_digest.hash}/{root_digest.size_bytes}:getTree" };
   }
+
+  // Split a blob into chunks.
+  //
+  // This splitting API aims to reduce download traffic between client and
+  // server, e.g., if a client needs to fetch a large blob that just has been
+  // modified slightly since the last built. In this case, there is no need to
+  // fetch the entire blob data, but just the binary differences between the two
+  // blob versions, which are typically determined by deduplication techniques
+  // such as content-defined chunking.
+  //
+  // Clients can use this API before downloading a blob to determine which parts
+  // of the blob are already present locally and do not need to be downloaded
+  // again. The server splits the blob into chunks according to a specified
+  // content-defined chunking algorithm and returns a list of the chunk digests
+  // in the order in which the chunks have to be concatenated to assemble the
+  // requested blob.
+  //
+  // A client can expect the following guarantees from the server if a split
+  // request is answered successfully:
+  //  1. The blob chunks are stored in CAS.
+  //  2. Concatenating the blob chunks in the order of the digest list returned
+  //     by the server results in the original blob.
+  //
+  // The usage of this API is optional for clients but it allows them to
+  // download only the missing parts of a large blob instead of the entire blob
+  // data, which in turn can considerably reduce download network traffic.
+  //
+  // Since the generated chunks are stored as blobs, they underlie the same
+  // lifetimes as other blobs. However, their lifetime is extended if they are
+  // part of the result of a split blob request.
+  //
+  // For the client, it is recommended to verify whether the digest of the blob
+  // assembled by the fetched chunks results in the requested blob digest.
+  //
+  // If several clients use blob splitting, it is recommended that they request
+  // the same splitting algorithm to benefit from each others chunking data. In
+  // combination with blob splicing, an agreement about the chunking algorithm
+  // is recommended since both client as well as server side can benefit from
+  // each others chunking data.
+  //
+  // Servers are free to implement this functionality, but they need to declare
+  // whether they support it or not by setting the
+  // [CacheCapabilities.blob_split_support][build.bazel.remote.execution.v2.CacheCapabilities.blob_split_support]
+  // field accordingly.
+  //
+  // Errors:
+  //
+  // * `NOT_FOUND`: The requested blob is not present in the CAS.
+  // * `RESOURCE_EXHAUSTED`: There is insufficient disk quota to store the blob
+  //   chunks.
+  rpc SplitBlob(SplitBlobRequest) returns (SplitBlobResponse) {
+    option (google.api.http) = { get: "/v2/{instance_name=**}/blobs/{blob_digest.hash}/{blob_digest.size_bytes}:splitBlob" };
+  }
+
+  // Splice a blob from chunks.
+  //
+  // This is the complementary operation to the
+  // [ContentAddressableStorage.SplitBlob][build.bazel.remote.execution.v2.ContentAddressableStorage.SplitBlob]
+  // function to handle the splitted upload of large blobs to save upload
+  // traffic.
+  //
+  // If a client needs to upload a large blob and is able to split a blob into
+  // chunks locally according to some content-defined chunking algorithm, it can
+  // first determine which parts of the blob are already available in the remote
+  // CAS and upload the missing chunks, and then use this API to instruct the
+  // server to splice the original blob from the remotely available blob chunks.
+  //
+  // In order to ensure data consistency of the CAS, the server will verify the
+  // spliced result whether digest calculation results in the provided digest
+  // from the request and will reject a splice request if this check fails.
+  //
+  // The usage of this API is optional for clients but it allows them to upload
+  // only the missing parts of a large blob instead of the entire blob data,
+  // which in turn can considerably reduce upload network traffic.
+  //
+  // In order to split a blob into chunks, it is recommended for the client to
+  // use one of the servers' advertised chunking algorithms by
+  // [CacheCapabilities.supported_chunking_algorithms][build.bazel.remote.execution.v2.CacheCapabilities.supported_chunking_algorithms]
+  // to benefit from each others chunking data. If several clients use blob
+  // splicing, it is recommended that they use the same splitting algorithm to
+  // split their blobs into chunks.
+  //
+  // Servers are free to implement this functionality, but they need to declare
+  // whether they support it or not by setting the
+  // [CacheCapabilities.blob_splice_support][build.bazel.remote.execution.v2.CacheCapabilities.blob_splice_support]
+  // field accordingly.
+  //
+  // Errors:
+  //
+  // * `NOT_FOUND`: At least one of the blob chunks is not present in the CAS.
+  // * `RESOURCE_EXHAUSTED`: There is insufficient disk quota to store the
+  //   spliced blob.
+  // * `INVALID_ARGUMENT`: The digest of the spliced blob is different from the
+  //   provided expected digest.
+  rpc SpliceBlob(SpliceBlobRequest) returns (SpliceBlobResponse) {
+    option (google.api.http) = { post: "/v2/{instance_name=**}/blobs:spliceBlob" body: "*" };
+  }
 }
 
 // The Capabilities service may be used by remote execution clients to query
@@ -1814,6 +1911,60 @@ message GetTreeResponse {
   string next_page_token = 2;
 }
 
+// A request message for
+// [ContentAddressableStorage.SplitBlob][build.bazel.remote.execution.v2.ContentAddressableStorage.SplitBlob].
+message SplitBlobRequest {
+  // The instance of the execution system to operate against. A server may
+  // support multiple instances of the execution system (with their own workers,
+  // storage, caches, etc.). The server MAY require use of this field to select
+  // between them in an implementation-defined fashion, otherwise it can be
+  // omitted.
+  string instance_name = 1;
+
+  // The digest of the blob to be splitted.
+  Digest blob_digest = 2;
+
+  // The chunking algorithm to be used. Must be IDENTITY (no chunking) or one of
+  // the algorithms advertised by the
+  // [CacheCapabilities.supported_chunking_algorithms][build.bazel.remote.execution.v2.CacheCapabilities.supported_chunking_algorithms]
+  // field.
+  ChunkingAlgorithm.Value chunking_algorithm = 3;
+}
+
+// A response message for
+// [ContentAddressableStorage.SplitBlob][build.bazel.remote.execution.v2.ContentAddressableStorage.SplitBlob].
+message SplitBlobResponse {
+  // The ordered list of digests of the chunks into which the blob was splitted.
+  // The original blob is assembled by concatenating the chunk data according to
+  // the order of the digests given by this list.
+  repeated Digest chunk_digests = 1;
+}
+
+// A request message for
+// [ContentAddressableStorage.SpliceBlob][build.bazel.remote.execution.v2.ContentAddressableStorage.SpliceBlob].
+message SpliceBlobRequest {
+  // The instance of the execution system to operate against. A server may
+  // support multiple instances of the execution system (with their own workers,
+  // storage, caches, etc.). The server MAY require use of this field to select
+  // between them in an implementation-defined fashion, otherwise it can be
+  // omitted.
+  string instance_name = 1;
+
+  // Expected digest of the spliced blob.
+  Digest blob_digest = 2;
+
+  // The ordered list of digests of the chunks which need to be concatenated to
+  // assemble the original blob.
+  repeated Digest chunk_digests = 3;
+}
+
+// A response message for
+// [ContentAddressableStorage.SpliceBlob][build.bazel.remote.execution.v2.ContentAddressableStorage.SpliceBlob].
+message SpliceBlobResponse {
+  // Computed digest of the spliced blob.
+  Digest blob_digest = 1;
+}
+
 // A request message for
 // [Capabilities.GetCapabilities][build.bazel.remote.execution.v2.Capabilities.GetCapabilities].
 message GetCapabilitiesRequest {
@@ -2000,6 +2151,40 @@ message Compressor {
   }
 }
 
+// Content-defined chunking algorithms used for splitting blobs into chunks.
+message ChunkingAlgorithm {
+  enum Value {
+    // No chunking. Servers MUST always support this, and do not need to
+    // advertise it.
+    IDENTITY = 0;
+
+    // Content-defined chunking using Rabin fingerprints. An implementation of
+    // this scheme in presented in this paper
+    // https://link.springer.com/chapter/10.1007/978-1-4613-9323-8_11. The final
+    // implementation of this algorithm should be configured to have the
+    // following properties on resulting chunk sizes.
+    //  - Minimum chunk size: 2 KB
+    //  - Average chunk size: 8 KB (0x00000000007FFFFF bit mask)
+    //  - Maximum chunk size: 64 KB
+    // The irreducible polynomial to be used for the modulo divisions is the
+    // following 64-bit polynomial of degree 53: 0x003DA3358B4DC173. The window
+    // size to be used is 64 bits.
+    RABINCDC = 1;
+
+    // Content-defined chunking using the FastCDC algorithm. The algorithm is
+    // described in this paper https://ieeexplore.ieee.org/document/9055082
+    // (Algorithm 2, FastCDC8KB). The algorithm is configured to have the
+    // following properties on resulting chunk sizes.
+    //  - Minimum chunk size: 2 KB
+    //  - Average chunk size: 8 KB
+    //  - Maximum chunk size: 64 KB
+    // The 256 64-bit random numbers in the Gear table are to be created with
+    // the Mersenne Twister pseudo-random number generator for 64-bit numbers
+    // with a state size of 19937 bits and a seed of 0.
+    FASTCDC = 2;
+  }
+}
+
 // Capabilities of the remote cache system.
 message CacheCapabilities {
   // All the digest functions supported by the remote cache.
@@ -2033,6 +2218,25 @@ message CacheCapabilities {
   // [BatchUpdateBlobs][build.bazel.remote.execution.v2.ContentAddressableStorage.BatchUpdateBlobs]
   // requests.
   repeated Compressor.Value supported_batch_update_compressors = 7;
+
+  // All the chunking algorithms supported by the remote cache. Remote cache may
+  // support multiple chunking algorithms simultaneously. Servers MUST support
+  // IDENTITY (no chunking), even if it is not listed here.
+  repeated ChunkingAlgorithm.Value supported_chunking_algorithms = 8;
+
+  // Whether blob splitting is supported for the particular server/instance. If
+  // yes, the server/instance implements the specified behavior for blob
+  // splitting and a meaningful result can be expected from the
+  // [ContentAddressableStorage.SplitBlob][build.bazel.remote.execution.v2.ContentAddressableStorage.SplitBlob]
+  // operation.
+  bool blob_split_support = 9;
+
+  // Whether blob splicing is supported for the particular server/instance. If
+  // yes, the server/instance implements the specified behavior for blob
+  // splicing and a meaningful result can be expected from the
+  // [ContentAddressableStorage.SpliceBlob][build.bazel.remote.execution.v2.ContentAddressableStorage.SpliceBlob]
+  // operation.
+  bool blob_splice_support = 10;
 }
 
 // Capabilities of the remote execution system.