An indexing pipeline builds indexes derived from source data. The index should always be converging to the current version of source data. In other words, once a new version of source data is processed by the pipeline, all data derived from previous versions should no longer exist in the target index storage. This is called data consistency requirement for an indexing pipeline. data consistency If you like our work, it would mean a lot to us if you could support Cocoindex on Github with a star. Thank you so much with a warm coconut hug 🥥🤗. support Cocoindex on Github with a star. Cocoindex on Github The Challenge of Concurrent Updates Compared to building single-process in-memory transformation logic, indexing pipeline has additional challenges in two aspects, to maintain data consistency: Temporally, it's long-running, i.e. the execution may span across multiple processes. e.g. an early execution may terminate in the middle, due to a SIGTERM, or a failure to commit to the target storage or a reboot. When the process restarts later, it should pick up existing states, and keep moving forward to update the target storage to desired status. This is both tedious and error-prone: for example, if a source data is only partially processed, and it gets updated before the restarted pipeline starts to process it again, how can we make sure states derived from the old version are properly cleared?


Spatially, execution happens concurrently, sometimes distributed across multiple machines for large workloads. This brings up the problem caused by out-of-order processing. Consider a scenario where your source data gets updated rapidly:

Version 1 (V1) exists in the system
Version 2 (V2) and Version 3 (V3) arrive within a second of each other
V2 takes longer to process than V3 Without proper handling, you might end up with V2 data in your index even though V3 is the latest version - effectively rolling back to stale data. Temporally, it's long-running, i.e. the execution may span across multiple processes. e.g. an early execution may terminate in the middle, due to a SIGTERM, or a failure to commit to the target storage or a reboot. When the process restarts later, it should pick up existing states, and keep moving forward to update the target storage to desired status. This is both tedious and error-prone: for example, if a source data is only partially processed, and it gets updated before the restarted pipeline starts to process it again, how can we make sure states derived from the old version are properly cleared? Temporally, it's long-running, i.e. the execution may span across multiple processes. e.g. an early execution may terminate in the middle, due to a SIGTERM, or a failure to commit to the target storage or a reboot. When the process restarts later, it should pick up existing states, and keep moving forward to update the target storage to desired status. This is both tedious and error-prone: for example, if a source data is only partially processed, and it gets updated before the restarted pipeline starts to process it again, how can we make sure states derived from the old version are properly cleared? Spatially, execution happens concurrently, sometimes distributed across multiple machines for large workloads. This brings up the problem caused by out-of-order processing. Consider a scenario where your source data gets updated rapidly:

Version 1 (V1) exists in the system
Version 2 (V2) and Version 3 (V3) arrive within a second of each other
V2 takes longer to process than V3 Without proper handling, you might end up with V2 data in your index even though V3 is the latest version - effectively rolling back to stale data. Spatially, execution happens concurrently, sometimes distributed across multiple machines for large workloads. This brings up the problem caused by out-of-order processing. Consider a scenario where your source data gets updated rapidly: Version 1 (V1) exists in the system
Version 2 (V2) and Version 3 (V3) arrive within a second of each other
V2 takes longer to process than V3 Without proper handling, you might end up with V2 data in your index even though V3 is the latest version - effectively rolling back to stale data. Version 1 (V1) exists in the system Version 2 (V2) and Version 3 (V3) arrive within a second of each other V2 takes longer to process than V3 Without proper handling, you might end up with V2 data in your index even though V3 is the latest version - effectively rolling back to stale data. Common Issues and Gaps Data Exposure Risks Data inconsistency in indexing systems can lead to several critical issues: System Integrity Stale data persists in indexes despite source updates
Inconsistent states across different storage systems
Mixed versions of the same source entry coexisting in the index Stale data persists in indexes despite source updates Inconsistent states across different storage systems Mixed versions of the same source entry coexisting in the index Security and Compliance Deleted sensitive information remains accessible through search interfaces
Outdated personal data violates privacy requirements (e.g., GDPR)
Regulatory non-compliance due to improperly retained records Deleted sensitive information remains accessible through search interfaces Outdated personal data violates privacy requirements (e.g., GDPR) Regulatory non-compliance due to improperly retained records AI System Reliability RAG systems incorporating outdated information into responses
Semantic search surfacing deprecated or retracted content
AI agents making decisions based on stale data RAG systems incorporating outdated information into responses Semantic search surfacing deprecated or retracted content AI agents making decisions based on stale data Business Operations Customer-facing applications showing incorrect information
Loss of trust due to inconsistent data presentation
Financial impact from outdated pricing or product details Customer-facing applications showing incorrect information Loss of trust due to inconsistent data presentation Financial impact from outdated pricing or product details Approaches for Maintaining Data Consistency 1. Tracking source-to-target Key Mappings If all your state management and storage systems are within a single transactional system (e.g., PostgreSQL), you can leverage database transactions for consistency. However, this is often not the case in real-world scenarios. For example, you may use an internal storage (e.g. PostgreSQL) to track per-source metadata, and uses an external vector store (e.g. Pinecone or Milvus) to store vector embeddings. There're extra challenges around this: We have to keep track of all the keys in target vector stores derived from each source key, in the internal storage. This is necessary when the old version of source data is updated or deleted: we need to locate previously derived keys and delete them if they no longer exist in the new version after processing.
However, the internal storage is not transactional with the external vector store, as they're two independent storage systems. So that if the process terminates abnormally in the middle, states in two systems will be out of sync. We have to keep track of all the keys in target vector stores derived from each source key, in the internal storage. This is necessary when the old version of source data is updated or deleted: we need to locate previously derived keys and delete them if they no longer exist in the new version after processing. However, the internal storage is not transactional with the external vector store, as they're two independent storage systems. So that if the process terminates abnormally in the middle, states in two systems will be out of sync. So we need to carefully design the order of write operations during commit, to avoid data leakage even when the process is only partially executed: Add new target keys to the internal store
Write to target storage

Create new records
Delete old records Note that for any target storage that supports atomic writing, this allows all updates derived from the same source key to be written atomically. This avoids mixed-version data on query time.


Remove no-longer-exist target keys from the internal store Add new target keys to the internal store Add new target keys to the internal store Write to target storage

Create new records
Delete old records Note that for any target storage that supports atomic writing, this allows all updates derived from the same source key to be written atomically. This avoids mixed-version data on query time. Write to target storage Create new records
Delete old records Note that for any target storage that supports atomic writing, this allows all updates derived from the same source key to be written atomically. This avoids mixed-version data on query time. Create new records Delete old records Note that for any target storage that supports atomic writing, this allows all updates derived from the same source key to be written atomically. This avoids mixed-version data on query time. Remove no-longer-exist target keys from the internal store Remove no-longer-exist target keys from the internal store This ensures a key invariant: keys tracked in the internal storage is always a superset of those really exist in the target store. This makes sure all the data in the target store is always tracked so never leaked. 2. Ordinal Bookkeeping Ordinals are unique identifiers that establish a strict ordering of data updates in your pipeline. They help bookkeep the sequence and timing of changes to ensure consistency. Common examples include: Timestamps (e.g., 2024-01-06T12:34:56Z) - Precise time-based ordering
Version numbers (e.g., v1.2.3, 20240106.1) - Sequential numbering Timestamps (e.g., 2024-01-06T12:34:56Z) - Precise time-based ordering Version numbers (e.g., v1.2.3, 20240106.1) - Sequential numbering Bookkeep ordinals carefully in two key stages: During ingestion: Record and preserve the source ordinal (e.g., timestamp or version) from the original data
During commit: Verify the source ordinal to prevent newer versions from being overwritten by older ones During ingestion: Record and preserve the source ordinal (e.g., timestamp or version) from the original data During commit: Verify the source ordinal to prevent newer versions from being overwritten by older ones Ordinal bookkeeping happens in Phase 1, which rejects out-of-order commits - they won't arise any updates in Phase 1, and Phase 2 and 3 won't be executed at all. 3. Versioned Unique Key Assignment with Soft Deletion When there's multiple ongoing processings happen on the same source key, updates writing to the target storage (Phase 2 above) may be out of order, which causes results derived from old versions overwriting newer ones. To avoid this, we need to: For data in the target storage, make sure the keys are unique even across different versions.
Leverage soft deletion in the target storage, i.e. deleted versions are represented by a row with a deleted field set. For data in the target storage, make sure the keys are unique even across different versions. Leverage soft deletion in the target storage, i.e. deleted versions are represented by a row with a deleted field set. deleted Specifically: In the transaction in Phase 1, we establish a full view of all derived keys for the current source key: collect existing keys, and generate new keys.
In Phase 2, we do UPSERT for both new and deleted versions - the only difference is the deleted field is set for deleted versions, and not touched for new versions.
During query time, we filter out all records with deleted field is set. In the transaction in Phase 1, we establish a full view of all derived keys for the current source key: collect existing keys, and generate new keys. In Phase 2, we do UPSERT for both new and deleted versions - the only difference is the deleted field is set for deleted versions, and not touched for new versions. deleted During query time, we filter out all records with deleted field is set. deleted Besides, we need an offline GC process to garbage collect rows in the target storage with the deleted field set. Because of this, the deleted field could be a timestamp of the deletion time, hence we can decide when to GC a specific version based on it. deleted deleted How we solved these at CocoIndex The discussions above only cover a part of the complexity of building and maintaining long-running or distributed indexing pipelines. There're other additional complexities, compared to building single-process in-memory transformation logic, such as: CRUD across different APIs
Incremental historical state management
Reuse existing computations
Data migration on processing logic changes CRUD across different APIs Incremental historical state management Reuse existing computations Data migration on processing logic changes CocoIndex framework aims at handling these complexities so users can focus on pure transformations: CocoIndex Users write transformations as pure functions

Output depends only on input
No side effects or hidden states
Easy to understand and maintain



Framework handles:

Previous computation result management
Incremental processing
State management
Data consistency
Data upgrade on processing logic changes Users write transformations as pure functions

Output depends only on input
No side effects or hidden states
Easy to understand and maintain Users write transformations as pure functions Output depends only on input
No side effects or hidden states
Easy to understand and maintain Output depends only on input No side effects or hidden states Easy to understand and maintain Framework handles:

Previous computation result management
Incremental processing
State management
Data consistency
Data upgrade on processing logic changes Framework handles: Previous computation result management
Incremental processing
State management
Data consistency
Data upgrade on processing logic changes Previous computation result management Incremental processing State management Data consistency Data upgrade on processing logic changes By exposing a data-driven programming paradigm, CocoIndex allows developers to focus on their core business logic while maintaining data consistency and freshness. CocoIndex If you like our work, Please support us - 🥥 Cocoindex on Github with a star. Thank you so much with a warm coconut hug 🥥🤗. Cocoindex on Github

This story contains new, firsthand information uncovered by the writer.

Stop Rebuilding Your Index From Scratch. There’s a Better Way.

Keep Your Indexes Fresh With This Real-time Pipeline

Star my work on github

Read My Stories

Too Long; Didn't Read

Get Free Unlimited IP Geo Data With IPinfo Lite

Your AI Assistant Might Be Using Stale Data. This Open Source Tool Wants to Fix That

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Build a Smarter Store: Let GPT Label Your Products and Predict What Sells Next

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

Build a Smarter Store: Let GPT Label Your Products and Predict What Sells Next

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps