Microsoft Technology Licensing, LLC

Naveen Prakash

Alejandro Hernandez Saenz

Cristian Diaconu

Hanumantha Rao Kodavalla

Alexander Budovski

Craig Steven Freedman

Panagiotis Antonopoulos

Krystyna Ewa Reisteter

Raghavendra Thallam Kodandaramaih

Daniel Georg Schall

Adrian-Leonard Radu

Rogério Ramos

Huanhui Hu

Sheetal Shrotri

Peter Byrne

Sandeep Lingam

Rogerio Ramos

Stefano Stefani

Cesar Alejandro Galindo-Legaria

Marc Todd Friedman

Milind Madhukar Joshi

Erik Ismert

Vikram Wakade

Marcel Van Der Holst

Distributed database systems including compute nodes and page servers are described herein that enable separating logical and physical storage of database files in a distributed database system. A distributed database system includes a page server and a compute node, and is configured to store a logical database file that includes data and is associated with a file identifier. Each page server is configurable to store slices (i.e., subportions) of the logical database file. The compute node is coupled to the plurality of page servers and configured to store the logical database file responsive to a received command. In an aspect, such storage may comprise slicing the data comprising the logical database file into a set of slices with each being associated with a respective page server, maintaining an endpoint mapping for each slice of the first set of slices, and transmitting each slice to the associated for storage thereby.

Separation of logical and physical storage in a distributed database system

Methods for operation fragmentation with metadata serialization in query processing pushdowns are performed by systems and devices. A compute node receives a query directed to database data, and generates query text fragments. Portions of metadata of the database are read from different page servers, and are serialized by the compute node. Page identities of data pages in a page server that stores the data are determined from a page index at the compute node, and the compute node provides the text fragments, the serialized metadata, and the page identities to the page server storing the data. The page server compiles the text fragments based on the serialized metadata to generate an executable query plan for the query. The page server initializes and performs execution of the executable query plan against the data as a pushdown query operation, and a result of pushdown query operation is provided to the compute node.

Operation fragmentation with metadata serialization in query processing pushdowns

Distributed database systems including compute nodes and page servers are described herein that enable compute nodes to pushdown certain query processing compute tasks to the page servers to take advantage of otherwise idle compute resources at the page servers, and to reduce the quantity of data that moves between compute nodes and page servers. A distributed database system includes a page server and a compute node, wherein the page server is configured to maintain multiple versions of stored data objects. The compute node is configured to receive a query and generate a transaction context (TC) and modified table schemas (MTS) scoped to the query, and pushdown the query, TC and MTS to the page server that is configured to determine which data objects at the page server satisfy the query, and for each such object, which version of the object should be returned based on the TC.

Schema and data modification concurrency in query processing pushdown

Methods for page split detection and affinity in query processing pushdowns are performed by systems and devices. Page servers perform pushdown operations based on specific, and specifically formatted or generated, information, instructions, and data provided thereto from a compute node. Page servers also determine that page splits have occurred during reading of data pages maintained by page servers during pushdown operations, and also during fulfillment of compute node data requests. To detect a data page has split, page servers utilize information from a compute node of an expected next data page which is compared to a next data page in the page server page index. A mismatch in the comparison by page servers indicates data page was split. Compute nodes and page servers store and maintain off-row data generated during data operations via page affinity considerations where the off-row data is stored at the same page server as the data.

Page split detection and affinity in query processing pushdowns

Distributed database systems including compute nodes and page servers are described herein that enable separating logical and physical storage of database files in a distributed database system. A distributed database system includes a page server and a compute node and is configured to store a logical database file that includes data and is associated with a file identifier. Each page server is configurable to store slices (i.e., subportions) of the logical database file. The compute node is coupled to the plurality of page servers and configured to store the logical database file responsive to a received command. In an aspect, such storage may comprise slicing the data comprising the logical database file into a set of slices with each being associated with a respective page server, maintaining an endpoint mapping for each slice of the first set of slices, and transmitting each slice to the associated for storage thereby.

Embodiments described herein detect data corruption in a distributed data set system. For example, a system comprises node(s) for processing queries with respect to a distributed data set comprising a plurality of storage segments. A write transaction resulting from a query with respect to a particular storage segment is logged in a log record that describes a modification to the storage segment. A log service provides the log record to a data server managing a portion of the distributed data set in which the storage segment is included, which performs the write transaction with respect to the storage segment. For redundancy purposes, the data server has replica(s) that manage respective replicas of the portion of the distributed data set managed thereby. For backup purposes, snapshots of the replica(s) are periodically generated. To determine a data corruption, a snapshot of one replica is cross-validated with a snapshot of another replica.

Snapshot-based data corruption detection

Brokering log records so as to prevent log records that are not yet persisted in a persistent log from being disseminated. The log records may be generated as a primary compute system performs operations. Upon receiving a request for a log record, the broker component determines whether the requested log record has been persisted in a persistent log. If the broker component determines that the log record has been persisted in the persistent log, the broker component responds to the request by causing the requested log record to be provided to the requesting entity (e.g., a secondary compute system). On the other hand, if the log record cannot yet determine that the log record has been persisted in the persistent log, the broker component prevents the log record from being provided to the requesting entity. This prevents data from being inconsistent during recovery.

Brokering persisted and unpersisted log records

Caching storage segments (e.g., pages) loaded from a remote storage such that, during recovery, the cached loaded storage segments may be at least partially recovered without reloading the storage segments from the remote storage. During normal operation of a computing system, storage segments are loaded from remote storage into local memory of a computing system. At some point, either due to eviction of the storage segment due to aging out of the storage segment, or due to writing of the storage segment, it is determined to write at least some of the loaded storage segments into local persistent storage. In conjunction with this, the corresponding storage segment is written to a respective storage address of the local persistent storage. Also, a correlation between an identifier of the storage segment and the respective address in the persistent storage is recorded in a persistent data structure.

Persistent storage segment caching for data recovery

The bifurcation of a log into a fixed-size log portion and a growable log portion. Log records are first written to the fixed-size log portion. At some point, the older log records within the fixed-size log portion are destaged to the growable log portion. Destaging may involve copying the log records from the fixed-size log portion and tacking the log record onto the end of the growable log portion. This destaging might occur one sub-portion (e.g., one virtual log file) at a time. The more recent log records are within the fixed-size log portion. Accordingly, new writes to the log occur to the fixed-size log portion. Furthermore, a large portion of the reads are of the more recent log records within the fixed-size log portion. Thus, optimizing performance within the fixed-size log portion results in significant improvement in the performance of the overall log.

Log destaging from fixed-size log portion

The use of log marking (otherwise known as 'coloring') of sub-portions of a log that records actions (e.g., data operations) performed by a computing system. The log is composed of multiple sub-portions, such as virtual log files, which are successively added to the log as the log grows. For instance, the sub-portions may be virtual log files of the log. The principles described herein change the use of log marking depending on which sub-portion of the log is being marked. If the computing system fails, and recovery is needed, the recovery process can thus deterministically identify where the last written log record is.

Log marking dependent on log sub-portion

Use of a log service computing system to apply a log to storage segments in the context of there being a primary compute system that operates upon data, and secondary compute system(s) that are at the ready to be promoted to the new primary compute system should the old primary compute system fail. Rather than have each of the primary and secondary compute systems maintain their own replicated of the entire data locally, the data is separated in storage segments that are collectively distributed amongst multiple storage segment servers.

Growing community of inventors

Sammamish, WA, United States of America

Naveen Prakash