Hilprecht, Benjamin (2022)
Data-Efficient Learned Database Components.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00022531
Dissertation, Erstveröffentlichung, Verlagsversion
Kurzbeschreibung (Abstract)
While databases are the backbone of many software systems, database components such as query optimizers often have to be redesigned to meet the increasing variety in workloads, data and hardware designs, which incurs significant engineering efforts to adapt their design. Recently, it was thus proposed to replace DBMS components such as optimizers, cardinality estimators, etc. by ML models, which not only eliminates the engineering efforts but also provides superior performance for many components.
The predominant approach to derive such learned components is workload-driven learning where ten thousands of queries have to be executed first to derive the necessary training data. Unfortunately, the training data collection, which can take days even for medium-sized datasets, has to be repeated for every new database (i.e., the combination of dataset, schema and workload) a component should be deployed for. This is especially problematic for cloud databases such as Snowflake or Redshift since this effort has to be incurred for every customer.
This dissertation thus proposes data-efficient learned database components, which either reduce or fully eliminate the high costs of training data collection for learned database components. In particular, three directions are proposed in this dissertation, namely (i) we first aim to reduce the number of training queries needed for workload-driven components before we (ii) propose data-driven learning, which uses the data stored in the database as training data instead of queries, and (iii) introduce zero-shot learned components, which can generalize to new databases out-of-the-box, s.t. no training data collection is required.
First, we strive to reduce the number of training queries required for workload-driven components by using simulation models to convey the basic tradeoffs of the underlying problem, e.g., that in database partitioning the network costs of shuffling tuples over the network for joins is the dominating factor. This substantially reduces the number of training queries since the basic principles are already covered by the simulation model and thus only subtleties not covered in the simulation model have to be learned by observing query executions, which we will demonstrate for the problem of database partitioning. An alternative direction is to incorporate domain knowledge (e.g., in a cost model we could encode that scan costs increase linearly with the number of tuples) into components by designing them using differentiable programming. This significantly reduces the number of learnable parameters and thus also the number of required training queries. We demonstrate the feasibility of the approach for the problem of cost estimation in databases. While both approaches reduce the number of training queries, there is still a significant number of training queries required for unseen databases.
This motivates our second approach of data-driven learning. In particular, we propose to train the database component by learning the data distribution present in a database instead of observing query executions. This not only completely eliminates the need to collect training data queries but can even improve the state-of-the-art in problems such as cardinality estimation or AQP. While we demonstrate the applicability to a wide range of additional database tasks such as the completion of incomplete relational datasets, data-driven learning is only useful for problems where the data distribution provides sufficient information for the underlying database task. However, for tasks where observations of query executions are indispensable such as cost estimation, data-driven learning cannot be leveraged.
In a third direction, we thus propose zero-shot learned database components, which are applicable to a broader set of tasks including those that require observations of queries. In particular, motivated by recent advances in transfer learning, we propose to pretrain a model once on a variety of databases and workloads and thus allow the component to generalize to unseen databases out-of-the-box. Hence, similar to data-driven learning no training queries have to be collected. In this dissertation, we demonstrate that zero-shot learning can indeed yield learned cost models which can predict query latencies on entirely unseen databases more accurately than state-of-the-art workload-driven approaches, which require ten thousands of query executions on every unseen database.
Overall, the proposed techniques yield state-of-the-art performance for many database tasks while significantly reducing or completely eliminating the expensive training data collection for unseen databases. However, while the proposed directions address the prevalent data-inefficiency of learned database components, there are still many opportunities to improve learned components in the future. First, the robustness and debuggability of learned components should be improved since as of today they do not offer the same transparency as standard code in databases, which can render the components less attractive to be deployed in production systems. Moreover, to increase the applicability of data-driven models it is desirable to increase the coverage of supported queries, e.g., queries involving wildcard predicates on string columns, which are currently not supported by data-driven learning. Finally, we envision that a broader set of tasks should be supported in the future by zero-shot models (e.g., query optimization) potentially converging towards complete zero-shot learned systems.
| Typ des Eintrags: | Dissertation | ||||
|---|---|---|---|---|---|
| Erschienen: | 2022 | ||||
| Autor(en): | Hilprecht, Benjamin | ||||
| Art des Eintrags: | Erstveröffentlichung | ||||
| Titel: | Data-Efficient Learned Database Components | ||||
| Sprache: | Englisch | ||||
| Referenten: | Binnig, Prof. Dr. Carsten ; Trummer, Prof. Dr. Immanuel | ||||
| Publikationsjahr: | 2022 | ||||
| Ort: | Darmstadt | ||||
| Kollation: | xxiii, 260 Seiten | ||||
| Datum der mündlichen Prüfung: | 18 Oktober 2022 | ||||
| DOI: | 10.26083/tuprints-00022531 | ||||
| URL / URN: | https://tuprints.ulb.tu-darmstadt.de/22531 | ||||
| Kurzbeschreibung (Abstract): | While databases are the backbone of many software systems, database components such as query optimizers often have to be redesigned to meet the increasing variety in workloads, data and hardware designs, which incurs significant engineering efforts to adapt their design. Recently, it was thus proposed to replace DBMS components such as optimizers, cardinality estimators, etc. by ML models, which not only eliminates the engineering efforts but also provides superior performance for many components. The predominant approach to derive such learned components is workload-driven learning where ten thousands of queries have to be executed first to derive the necessary training data. Unfortunately, the training data collection, which can take days even for medium-sized datasets, has to be repeated for every new database (i.e., the combination of dataset, schema and workload) a component should be deployed for. This is especially problematic for cloud databases such as Snowflake or Redshift since this effort has to be incurred for every customer. This dissertation thus proposes data-efficient learned database components, which either reduce or fully eliminate the high costs of training data collection for learned database components. In particular, three directions are proposed in this dissertation, namely (i) we first aim to reduce the number of training queries needed for workload-driven components before we (ii) propose data-driven learning, which uses the data stored in the database as training data instead of queries, and (iii) introduce zero-shot learned components, which can generalize to new databases out-of-the-box, s.t. no training data collection is required. First, we strive to reduce the number of training queries required for workload-driven components by using simulation models to convey the basic tradeoffs of the underlying problem, e.g., that in database partitioning the network costs of shuffling tuples over the network for joins is the dominating factor. This substantially reduces the number of training queries since the basic principles are already covered by the simulation model and thus only subtleties not covered in the simulation model have to be learned by observing query executions, which we will demonstrate for the problem of database partitioning. An alternative direction is to incorporate domain knowledge (e.g., in a cost model we could encode that scan costs increase linearly with the number of tuples) into components by designing them using differentiable programming. This significantly reduces the number of learnable parameters and thus also the number of required training queries. We demonstrate the feasibility of the approach for the problem of cost estimation in databases. While both approaches reduce the number of training queries, there is still a significant number of training queries required for unseen databases. This motivates our second approach of data-driven learning. In particular, we propose to train the database component by learning the data distribution present in a database instead of observing query executions. This not only completely eliminates the need to collect training data queries but can even improve the state-of-the-art in problems such as cardinality estimation or AQP. While we demonstrate the applicability to a wide range of additional database tasks such as the completion of incomplete relational datasets, data-driven learning is only useful for problems where the data distribution provides sufficient information for the underlying database task. However, for tasks where observations of query executions are indispensable such as cost estimation, data-driven learning cannot be leveraged. In a third direction, we thus propose zero-shot learned database components, which are applicable to a broader set of tasks including those that require observations of queries. In particular, motivated by recent advances in transfer learning, we propose to pretrain a model once on a variety of databases and workloads and thus allow the component to generalize to unseen databases out-of-the-box. Hence, similar to data-driven learning no training queries have to be collected. In this dissertation, we demonstrate that zero-shot learning can indeed yield learned cost models which can predict query latencies on entirely unseen databases more accurately than state-of-the-art workload-driven approaches, which require ten thousands of query executions on every unseen database. Overall, the proposed techniques yield state-of-the-art performance for many database tasks while significantly reducing or completely eliminating the expensive training data collection for unseen databases. However, while the proposed directions address the prevalent data-inefficiency of learned database components, there are still many opportunities to improve learned components in the future. First, the robustness and debuggability of learned components should be improved since as of today they do not offer the same transparency as standard code in databases, which can render the components less attractive to be deployed in production systems. Moreover, to increase the applicability of data-driven models it is desirable to increase the coverage of supported queries, e.g., queries involving wildcard predicates on string columns, which are currently not supported by data-driven learning. Finally, we envision that a broader set of tasks should be supported in the future by zero-shot models (e.g., query optimization) potentially converging towards complete zero-shot learned systems. |
||||
| Alternatives oder übersetztes Abstract: |
|
||||
| Status: | Verlagsversion | ||||
| URN: | urn:nbn:de:tuda-tuprints-225311 | ||||
| Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik | ||||
| Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Data Management (2022 umbenannt in Data and AI Systems) |
||||
| Hinterlegungsdatum: | 21 Okt 2022 12:02 | ||||
| Letzte Änderung: | 24 Okt 2022 07:17 | ||||
| PPN: | |||||
| Referenten: | Binnig, Prof. Dr. Carsten ; Trummer, Prof. Dr. Immanuel | ||||
| Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 18 Oktober 2022 | ||||
| Export: | |||||
| Suche nach Titel in: | TUfind oder in Google | ||||
 |
Frage zum Eintrag |
Optionen (nur für Redakteure)
 |
Redaktionelle Details anzeigen |
Drucken
Impressum