Jan 2026
AI 시대의 CUBRID
CUBVEC — 벡터 검색 · 사내 LLM · OLTP에서 분석까지
AI · LLM 환경에서 DBMS가 새로 떠안게 된 역할 — 특히 벡터 검색 — 을 CUBRID가 어떻게 풀고 있는지 정리. CUBVEC 프로젝트의 인덱스 구조, 임베딩 모델과의 통합, 사내(Private) LLM이 요구하는 데이터 파이프라인, OLTP를 넘어 분석 워크로드로 확장하는 방향을 다룬다.
Hyung-Gyu Ryoo's
Talks & Notes
주로 발표 자료를 모아둡니다. 곁가지로 코드 분석 노트와 읽은 책 · 논문에 대한 메모도 함께 둡니다. 관심사는 데이터베이스 내부와 그 주변의 시스템 소프트웨어입니다.
함께 보기 · Notes & Analyses ↗
Featured
Talks
May 2026
CUBRID Lock Manager draft
Multi-Granularity Locking, Conversion, Deadlock Detection
How the textbook 2PL and multi-granularity locking get realized on top of CUBRID's OID-shaped objects. Walks through the core data structures (LK_ENTRY, LK_RES, the aggregate-mode cache), the 12-mode compatibility matrix, the acquisition flow in lock_object, how the release path is scoped by MVCC-disabled class set and isolation level (with X locks always commit-bound), and the waits-for-graph deadlock detector with its most-recently-blocked victim policy.
docs Code analysis link
May 2026
CUBRID Architecture Overview draft
Processes, Layered Stack, Query Pipeline, Distribution
Front-door router for the CUBRID code-analysis tree. Names the four long-lived processes (cub_master, cub_server, cub_pl, cub_broker + cub_cas) and their IPC; walks the layered storage stack from disk to workspace; traces the query pipeline from parser to cursor; sketches the concurrency axis (MVCC + lock + log + checkpoint + recovery + DWB), the distribution layer (heartbeat + HA + CDC + 2PC + flashback + backup), the PL family, and the cross-cutting infrastructure. One diagram per axis, with cross-refs into ~70 detail docs across eight subcategories.
docs Code analysis link
May 2026
CUBRID Locator draft
OID Workspace, Bulk Fetch/Flush, Server-Side Insert/Update/Delete Bridge
How CUBRID bridges in-memory objects and on-disk OIDs. A client-side workspace of MOPs marshals dirty objects into bidirectional LC_COPYAREA buffers (LC_COPYAREA_MANYOBJS + LC_COPYAREA_ONEOBJ records); locator_mflush_cache batches them at flush time; the server-side locator_attribute_info_force / locator_{insert,update,delete}_force family is the one canonical entry point that fans out into heap, btree, lock, log, FK, and replication.
docs Code analysis link
May 2026
CUBRID MVCC draft
Snapshot Construction, Active-MVCCID Tracking, Vacuum Coordination
How CUBRID implements snapshot isolation on top of a lazily-issued 64-bit MVCCID, a bit-array-plus-long-tran active set (`mvcc_active_tran`), and a 2048-slot history ring (`mvcctable`). Walks through the per-record `mvcc_rec_header`, the lock-free snapshot construction in `build_mvcc_info`, the three-valued `mvcc_satisfies_snapshot` predicate, and how `complete_mvcc` advances the ring and the `m_oldest_visible` watermark that gates vacuum.
docs Code analysis link
May 2026
CUBRID Query Processing draft
Parser, Optimizer, XASL, Executor, Access Methods
How CUBRID turns a SQL string into a stream of tuples. Flex+Bison parses into a PT_NODE tree, semantic check resolves names and types and pushes predicates to CNF, the rewriter flattens subqueries / inlines views / lowers LIMIT, a DP-style join enumerator with a System R cost model emits a QO_PLAN that the XGEN walk lowers to an XASL_NODE tree, a SHA-1 keyed XASL cache short-circuits compile work on re-execute, the Volcano-style executor drives open/next/close over SCAN_ID handles that dispatch across heap / btree / list / value / json-table / show / dblink access methods, predicate evaluation reuses one PRED_EXPR walker for scan filters and joins, and every materialised tuple stream lands on the QFILE_LIST_ID list-file substrate that overflows to FILE_TEMP before the cursor hands rows back through the broker / CCI.
docs Code analysis link
May 2026
CUBRID Storage Engine draft
Disks, Pages, Heaps, B+Trees, Hash, Overflow
A layered tour of the CUBRID storage engine: disk_manager owns volumes, sectors (64-page allocation units), files, and pages; the page buffer caches pages through a per-bucket hash and a three-zone LRU split into per-thread private lists with adjustable quotas plus a shared list; the double-write buffer stages every dirty page sequentially before the home write to defeat torn writes; heap_manager, btree, and extendible_hash sit on top of the buffer, never on the disk. WAL ordering is structural — the buffer refuses to flush a dirty page until the matching log LSA is durable.
docs Code analysis link
May 2026
CUBRID Transaction & Recovery draft
MVCC + Locks, WAL, ARIES, Vacuum, 2PC
How CUBRID realises ACID across eleven modules: MVCC and the lock manager carry isolation, while the log manager + prior list + checkpoint + recovery manager (with the double-write buffer as a Storage-Engine cross-section) underwrite atomicity and durability. Three extensions — 2PC, flashback, backup/restore — reuse the same TDES and WAL machinery for cross-server commit, time-travel queries, and point-in-time recovery.
docs Code analysis link
Sep 2025
DiskANN and Related Works on Disk-Based k-ANN Indexes
CUBRID 디스크 기반 k-ANN 인덱스 구현을 위한 논문 리뷰
CUBRID의 디스크 기반 k-ANN 인덱스 구현 준비를 위한 사내 논문 리뷰. DiskANN (NeurIPS 2019), AiSAQ (arXiv 2024), FreshDiskANN (arXiv 2021) 세 편을 차례로 정리하고, 각 알고리즘이 SSD 라운드트립을 최소화하면서 십억 단위 벡터를 검색해내는 방식 — Vamana 그래프, Product Quantization, 빔 검색, 스트리밍 인덱싱 — 을 비교한다.
May 2025
Survey of Vector Database Management Systems
Paper review: Pan · Wang · Li (VLDBJ 2024)
사내 paper seminar 발표. Pan·Wang·Li의 VDBMS 서베이 (VLDB Journal 2024 / arXiv:2310.14021) 를 8개 섹션 (Introduction · Query Processing · Indexing · Query Optimization and Execution · Current Systems · Benchmarks · Challenges and Open Problem · Conclusion) 로 풀어 정리. CUBRID의 벡터 검색 로드맵(CUBVEC) 을 염두에 두고 인덱싱·쿼리 최적화·현행 시스템 비교를 훑는 reading 세션 슬라이드. 2026-05 보강 — PowerPoint 원본 43 페이지에 챕터 입구 overview · Benchmarks · Challenges and Open Problems 16 페이지를 Marp 로 작성해 PDF 병합 (총 59 페이지).
Nov 2021
오픈소스에서 개발자와 함께 성장하기
CUBRID 기여 튜토리얼 — 프로젝트 선택, 기여 진입점, 개발 프로세스
공개SW페스티벌 2021의 두 번째 세션. 'OPEN COMMUNITY' 라는 CUBRID의 지향점을 출발점으로, 외부 기여자가 처음 어떤 프로젝트를 고르고, 어디서 진입점(문서 갭, IN PROGRESS 이슈, 빌드 셋업, 단순 픽스)을 찾고, JIRA · GitHub · 코드리뷰 · 자동 테스트로 이어지는 흐름을 직접 거쳐온 경험으로 안내한다. CUBRID Foundation (2020-02 시애틀 설립), CUBRID 11.0의 라이선스 전환 (BSD → Apache 2.0 클라이언트 / 코어), 그리고 모듈별 저장소 구조까지 다룬다.
Sep 2021
CUBRID 오픈소스 개발 프로세스
MySQL / MariaDB / PostgreSQL와의 비교, CUBRID 4단계 (2008–2019), 그리고 OSS DBMS 비즈모델
DD튜브 발표. 대표적인 오픈소스 RDBMS — MySQL · MariaDB · PostgreSQL · CUBRID — 의 개발 히스토리를 비교하고, CUBRID가 2008년 오픈소스 전환 이후 거쳐온 네 단계를 정리. OSS 라이선스 선택, G-Cloud · D-Cloud 같은 공공 클라우드 채택 사례, 그리고 오픈소스 DBMS 비즈니스 모델 — '코드를 무료로 풀면 무엇으로 먹고사는가' — 까지 다룬다.
Nov 2020
오픈소스 데이터베이스, 큐브리드에 기여해보기
KCD 2020 발표 · JIRA · GitHub · PR · CI
KCD (Korea Community Day) 2020에서 한 발표. 오픈소스 RDBMS인 CUBRID에 외부 기여자로서 어떻게 합류하는지를 다섯 단계 (Communication · Open Issue · Implementation · Pull Request · Testing) 로 풀고, JIRA 이슈 카테고리, GitHub PR 코드리뷰 흐름, 그리고 QA 팀이 돌리는 네 가지 테스트 스위트 (medium · SQL · performance · heavy) 까지 안내한다.
Dec 2018
Working with 3D Data in PostGIS via SFCGAL
3D extrusion, boolean ops, and what PostGIS still can't do (as of 2018)
Walks through PostGIS's SFCGAL-backed 3D toolbox: ST_Extrude turning NYC building footprints into 3D buildings, visualizing them in QGIS 3.4's 3D map view, and SFCGAL boolean / intersection / union / difference on solids. Closes with the practical limits of using PostGIS as a 3D-data service backend.
Aug 2017
Development of an Extension of GeoServer for Handling 3D Spatial Data
FOSS4G 2017 · Boston
Talk given at FOSS4G 2017 (Boston) on extending GeoServer and GeoTools so that the Web Feature Service stack can carry 3D geometries — covering the modifications to GeoTools' geometry model and the WFS handlers in GeoServer.
Reading
May 2026
CUBRID Lock Manager draft
Multi-Granularity Locking, Conversion, Deadlock Detection
How the textbook 2PL and multi-granularity locking get realized on top of CUBRID's OID-shaped objects. Walks through the core data structures (LK_ENTRY, LK_RES, the aggregate-mode cache), the 12-mode compatibility matrix, the acquisition flow in lock_object, how the release path is scoped by MVCC-disabled class set and isolation level (with X locks always commit-bound), and the waits-for-graph deadlock detector with its most-recently-blocked victim policy.
docs Code analysis link
May 2026
CUBRID Architecture Overview draft
Processes, Layered Stack, Query Pipeline, Distribution
Front-door router for the CUBRID code-analysis tree. Names the four long-lived processes (cub_master, cub_server, cub_pl, cub_broker + cub_cas) and their IPC; walks the layered storage stack from disk to workspace; traces the query pipeline from parser to cursor; sketches the concurrency axis (MVCC + lock + log + checkpoint + recovery + DWB), the distribution layer (heartbeat + HA + CDC + 2PC + flashback + backup), the PL family, and the cross-cutting infrastructure. One diagram per axis, with cross-refs into ~70 detail docs across eight subcategories.
docs Code analysis link
May 2026
CUBRID Locator draft
OID Workspace, Bulk Fetch/Flush, Server-Side Insert/Update/Delete Bridge
How CUBRID bridges in-memory objects and on-disk OIDs. A client-side workspace of MOPs marshals dirty objects into bidirectional LC_COPYAREA buffers (LC_COPYAREA_MANYOBJS + LC_COPYAREA_ONEOBJ records); locator_mflush_cache batches them at flush time; the server-side locator_attribute_info_force / locator_{insert,update,delete}_force family is the one canonical entry point that fans out into heap, btree, lock, log, FK, and replication.
docs Code analysis link
May 2026
CUBRID MVCC draft
Snapshot Construction, Active-MVCCID Tracking, Vacuum Coordination
How CUBRID implements snapshot isolation on top of a lazily-issued 64-bit MVCCID, a bit-array-plus-long-tran active set (`mvcc_active_tran`), and a 2048-slot history ring (`mvcctable`). Walks through the per-record `mvcc_rec_header`, the lock-free snapshot construction in `build_mvcc_info`, the three-valued `mvcc_satisfies_snapshot` predicate, and how `complete_mvcc` advances the ring and the `m_oldest_visible` watermark that gates vacuum.
docs Code analysis link
May 2026
CUBRID Query Processing draft
Parser, Optimizer, XASL, Executor, Access Methods
How CUBRID turns a SQL string into a stream of tuples. Flex+Bison parses into a PT_NODE tree, semantic check resolves names and types and pushes predicates to CNF, the rewriter flattens subqueries / inlines views / lowers LIMIT, a DP-style join enumerator with a System R cost model emits a QO_PLAN that the XGEN walk lowers to an XASL_NODE tree, a SHA-1 keyed XASL cache short-circuits compile work on re-execute, the Volcano-style executor drives open/next/close over SCAN_ID handles that dispatch across heap / btree / list / value / json-table / show / dblink access methods, predicate evaluation reuses one PRED_EXPR walker for scan filters and joins, and every materialised tuple stream lands on the QFILE_LIST_ID list-file substrate that overflows to FILE_TEMP before the cursor hands rows back through the broker / CCI.
docs Code analysis link
May 2026
CUBRID Storage Engine draft
Disks, Pages, Heaps, B+Trees, Hash, Overflow
A layered tour of the CUBRID storage engine: disk_manager owns volumes, sectors (64-page allocation units), files, and pages; the page buffer caches pages through a per-bucket hash and a three-zone LRU split into per-thread private lists with adjustable quotas plus a shared list; the double-write buffer stages every dirty page sequentially before the home write to defeat torn writes; heap_manager, btree, and extendible_hash sit on top of the buffer, never on the disk. WAL ordering is structural — the buffer refuses to flush a dirty page until the matching log LSA is durable.
docs Code analysis link
May 2026
CUBRID Transaction & Recovery draft
MVCC + Locks, WAL, ARIES, Vacuum, 2PC
How CUBRID realises ACID across eleven modules: MVCC and the lock manager carry isolation, while the log manager + prior list + checkpoint + recovery manager (with the double-write buffer as a Storage-Engine cross-section) underwrite atomicity and durability. Three extensions — 2PC, flashback, backup/restore — reuse the same TDES and WAL machinery for cross-server commit, time-travel queries, and point-in-time recovery.
docs Code analysis link