Fichas
Data (All)
Fichas del conjunto (30)
- 1What is distributed data storagedistributed data storage means storing data across multiple machines or nodes so it can be accessed, processed, and scaled as one logical system. It is used to handle large datasets, improve fault tolerance, and support faster parallel processing in ML workflows
- 2What is data parallelismSplitting a dataset across multiple devices
- 3What is model parallelismPartitioning a model across multiple devices
- 4What are some considerations of data parallelism?Each node hosts a copy of the model and operations can be run synchronously or asynchronously (synchronous is common for deep learning). As each node has a copy of the model the hardware must have enough memory to host it.
- 5What is federated learning?Multiple nodes train copies of a model on subsets of data, but the data partitions are not created from a centralised data pool. Model updates are aggregated between nodes.
- 6What are the benefits of distributed ML?1. Efficiency (tasks can be completed fast)
2. Scalability (does not require significant infra investment)
3. Redundancy (fault tolerance - the ability to continue operating even if individual nodes fail - replicating processes preserves uptime so that failures don't interrupt the service) - 7Simply what are the 3 main challenges of distributed ML implementations?1. Network bottlenecks
2. Synchronisation overhead
3. Energy consumption - 8In distributed systems what are network bottlenecksInsufficient bandwidth between nodes is a frequent cause of bottlenecks in distributed ML networks. Methods such as tensor parallelism that require more communication are the most bandwidth-hungry. If the network is unable to provide the required bandwidth, distributed ML projects will face increased training times and reduced scalability.
- 9In distributed systems what is synchronisation overheadSynchronization overhead is a delay that occurs when one task cannot begin until another task is complete. In a synchronous system, all worker nodes must upload shared data before moving to the next stage of training. This moment is known as a synchronization barrier because the next phase does not begin until all nodes are synchronized. Straggler nodes can slow the process.
- 10What contributed to energy consumption in distributed systems- Computational requirements (GPUS)
- Communication (high speed networks)
- Cloud or edge computing (data centers)
- Algorithm and data processing choices (model efficiency - algorithms, feature engineering) - 11What are the 4 core ideas of distributed data storage?1. Partitioning (data is split into chunks and spread across nodes)
2. Replication (Copies are stored on multiple nodes to improve availability and durability)
3. Horizontal scaling (Add more machines)
4. Fault tolerance (If one node fails the system can keep working from other replicas) - 12What are some examples of distributed storage systemsApache Cassandra, Amazon DynamoDB, Google Bigtable, AWS S3
- 13What is a relational database?A relational database is a type of database that organizes data into rows and columns, which collectively form a table where the data points are related to each other.
Data is typically structured across multiple tables, which can be joined together via a primary key or a foreign key. These unique identifiers demonstrate the different relationships which exist between tables, and these relationships are usually illustrated through different types of data models. - 14What is a NoSQL database?NoSQL, also referred to as “not only SQL” or “non-SQL”, is an approach to database design that enables the storage and querying of data outside the traditional structures found in relational databases. Instead of the typical tabular structure of a relational database, NoSQL databases, house data within one data structure, such as JSON document. Since this non-relational database design does not require a schema, it offers rapid scalability to manage large and typically unstructured data sets.
NoSQL is also type of distributed database, which means that information is copied and stored on various servers, which can be remote or local. This ensures availability and reliability of data. If some of the data goes offline, the rest of the database can continue to run. - 15What is postgreSQLPostgreSQL is a classic relational database. It is strong for application backends, transactions, data integrity, foreign keys, joins, and complex querying
- 16What is MongoDBMongoDB is a document database in the NoSQL family. It stores data as JSON-like documents, which makes it flexible when your schema changes often or your records vary a lot
- 17What is BigQuery?A cloud data warehouse for analytics. It is used for huge scans, aggregations, dashboards, and BI rather than low-latency transactional reads and writes.
- 18What is Neo4jA graph based database
- 19What is a feature store?A feature store is a centralized repository for storing, sharing, and managing features used in machine learning models. It acts as a bridge between raw data and model training/deployment, ensuring that features are consistently defined, versioned, and accessible across different teams and projects.
- 20What are the 4 common storage patterns for a feature store?Offline-only, online-only, hybrid pattern, streaming pattern
- 21Simply - what is an online feature store?Primarily designed to serve real-time features to production models. These feature stores enable low-latency access, ensuring that models can respond quickly to incoming data.
- 22Simply what is an offline feature storeUsed primarily for batch processing, these stores are optimized for training and validating models. They often handle large volumes of historical data and are not concerned with real-time performance. where throughput and historical correctness matter more than latency. Offline stores solve training data generation, backfills, and point-in-time correct datasets for model training
- 23What are the key characteristics of an online feature store?1. Low latency (quick access, high performing DBs)
2. Real time data ingestion (handle data streams, allowing for real time feature computation)
3. Scalability (Horizontal scaling to accommodate varying loads of requests)
4. Consistency and freshness - 24What are the key characteristics of an offline feature store?1. Batch processing (working with large datasets)
2. Historical data access (analysis of trends and features that require long term patterns)
3. Data transformation (support aggregates to enhance quality of features)
4. Version control (track changes to features over time) - 25What is the typical process for a prediction service accessing streamed data?1. Producers send events like clicks, transactions or sensor updates into Kafka or Kinesis
2. A stream processor such as Fink or Spark computes rolling features, joins, or filters
3. Features are written to a low-latency store like redis, cassandra, dynamoDB
4. The prediction services looks up the latest features by entity key - 26What is direct and indirect access for a model service accessing features?- Indirect access is the most common and safest design: the model service queries a feature store or key-value store, which keeps latency low and isolates the model from stream complexity
- Direct access means the model service itself consumes the stream and keeps local state, but this is harder to operate and usually reserved for special low-latency case - 27What are benefits of using a feature store?1. Reusability and efficiency (Features are only computed once)
2. Standardisation and governance (single source of truth, helps compliance)
3. Consistency between training and serving
4. Low-latency access for real-time (optimised online store)
5. Integration of batch and streaming data - 28What is object storage suited for?Large, unstructured, or binary data. Model weights + checkpoints, training datasets (multimodal), preprocessed feature datasets, experimental artifacts, serialised models, versioning artifacts
- 29What is database storage suited for?Metadata, relationships, queryability, experiment metadata, hyperparams, model registry records, feature store metadata
- 30What is ETL?Extract, transform and load. Moving data from source into a destination
Información
- Descripción
- Software Engineer ML Theory Fichas sobre Data (All), creado por Sophie Cooper el 13/04/2026.
- Creado por
- Sophie Cooper
- Visitas
- 1
- Valoración