Research map / Architectures and efficiency / Pruning and efficient tuning
In-Context Adaptation to Concept Drift for Learned Database Operations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FLAIR, an online in-context adaptation framework for learned database operations that uses a task featurizer plus a PFN-style dynamic decision engine meta-trained on synthetic Bayesian priors, adapting at inference via a FIFO context memory of recent queries and true results instead of fine-tuning.
Problem
Concept drift in dynamic databases (frequent inserts, deletes, updates) degrades learned database operations such as cardinality estimation, and existing reactive retraining/fine-tuning approaches are costly and delayed.
Approach
Treat adaptation as conditional prediction f:(x | C_t) -> y, where a meta-trained prior-data fitted network (PFN-style DDE) conditions on a FIFO context memory of recent query–execution-result pairs, so no runtime parameter optimization is needed; a cross-attention task featurization module encodes database and query into a task vector.
Key results
- On JOB-light under mild drift, FLAIR achieves GMQ 2.36 versus 9.36 for PostgreSQL and 4.16 for DDUp.
- FLAIR accelerates adaptation by 5.2x and reduces GMQ by 22.5% for cardinality estimation compared with the best baseline.
- Integrated into PostgreSQL, FLAIR speeds up query execution by up to 1.9x over PostgreSQL's built-in estimator.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FLAIR instead of Fine-tune | Cardinality estimation · STATS, mild drift | GMQ: 5.35 -> 4.49 | Helped |
| Use FLAIR instead of Fine-tune | Cardinality estimation · STATS, severe drift | GMQ: 5.02 -> 5.47 | Hurt |
| Use FLAIR instead of Fine-tune | Cardinality estimation · Job-light, mild drift | GMQ: 2.45 -> 2.36 | Helped |
| Use FLAIR instead of Fine-tune | Cardinality estimation · Job-light, severe drift | GMQ: 8.09 -> 7.95 | Helped |
| Use FLAIR instead of PostgreSQL | Cardinality estimation · STATS, mild drift | GMQ: 174.38 -> 4.49 | Helped |
| Use FLAIR instead of PostgreSQL | Cardinality estimation · STATS, severe drift | GMQ: 293.47 -> 5.47 | Helped |
| Use FLAIR instead of PostgreSQL | Cardinality estimation · Job-light, mild drift | GMQ: 9.36 -> 2.36 | Helped |
| Use FLAIR instead of PostgreSQL | Cardinality estimation · Job-light, severe drift | GMQ: 32.09 -> 7.95 | Helped |
| Use in-context adaptation with dynamic context memory instead of reactive retraining or distillation | FLAIR · STATS and Job-light, long-term incremental drift | error accumulation | Helped |
| Enable embedding caching to avoid recomputation on repeated inputs | FLAIR · Inference | inference time | Helped |
Surprising findings
- Knowledge distillation (DDUp) underperforms fine-tuning under long-term gradual drift because it can replicate past errors.
Limitations
- Storage overhead is slightly higher than baselines.
- Theoretical error bound assumes i.i.d. data from a continuous distribution and depends on context size.
Experimental setup
Datasets: STATS JOB-light IMDB
Benchmarks: STATS JOB-light
Metrics: GMQ Q-error P-error MRE accuracy F1 MSE R^2 adaptation time
Baselines: PostgreSQL DeepDB ALECE DDUp Fine-tune DBest++ AutoGluon XGBoost LightGBM CatBoost
Builds on: prior-data fitted networks in-context learning NeuroCard DDUp
Closest papers by meaning
- Beyond Static Pipelines: Learning Dynamic Workflows for Text-to-SQL (ICML 2026 · similarity 0.85)
- Let's (not) just put things in Context: Test-time Training for Long-context LLMs (ICLR 2026 · similarity 0.83)
- Adaptive Data Optimization: Dynamic Sample Selection with Scaling Laws (ICLR 2025 · similarity 0.83)
- A Pre-training Framework for Relational Data with Information-theoretic Principles (NeurIPS 2025 · similarity 0.83)
- Theoretical Analysis of Learned Database Operations under Distribution Shift through Distribution Learnability (ICML 2024 · similarity 0.83)
- Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.