Performance Tuning for CDC: Managing Replication Lag in Azure Database for PostgreSQL with Debezium
Written By: Shashikant Shakya, Ashutosh Bapat, and Guangnan Shi The Problem Picture this: your CDC pipeline is running smoothly, streaming changes from PostgreSQL to Kafka. Then, a bulk update hits millions of rows. Suddenly, Kafka queues pile up, downstream systems lag; dashboards go stale. Why does replication lag spike during heavy operations? And what can you do about it? Why This Matters Change Data Capture (CDC) powers real-time integrations, pushing row-level changes from OLTP systems into event streams, data lakes, caches, and microservices. Debezium is a leading open-source CDC engine for PostgreSQL, and many teams successfully run Debezium against Azure Database for PostgreSQL to keep downstream systems synchronized. However, during large DML operations (bulk updates, deletes) or schema changes (DDL), replication lag can occur because: Debezium consumes WAL slower than the database produces it Kafka throughput dips Consumers fall behind This article explains why lag happens, grounded in logical decoding internals, and shows how to diagnose it quickly and what to tune across the database, Azure, and connector layers to keep pipelines healthy under heavy load. CDC Basics CDC streams incremental changes (INSERT/UPDATE/DELETE) from your source database to downstream systems in near real-time. In PostgreSQL, CDC is typically implemented using logical decoding and logical replication: PostgreSQL records every change in the Write-Ahead Log (WAL) WALSender reads WAL and decodes it into change events The pgoutput extension formats those changes, while Debezium subscribes and publishes them to Kafka topics Benefits of CDC: Low latency Lower source overhead than periodic full extracts Preserves transactional ordering for consumers The Internals: Why Lag Happens Replication lag during heavy operations isn’t random, it’s rooted in how PostgreSQL handles logical decoding. To understand why, let’s look at the components that process changes and what happens when they hit resource limits. Logical Decoding & ReorderBuffer Logical decoding reconstructs transaction-level changes so they can be delivered in commit order. The core component enabling this is the ReorderBuffer. What ReorderBuffer does: Reads WAL and groups changes per transaction, keeping them in memory until commit If memory exceeds logical_decoding_work_mem , PostgreSQL spills decoded changes to disk in per-slot spill files On commit, it reads back spilled data and emits changes to the client (via pgoutput → Debezium) Disk Spill Mechanics (Deep Dive) When a transaction is too large for memory: PostgreSQL writes decoded changes to spill files under pg_replslot/<slot_name>/ Wait events like ReorderBufferWrite and ReorderBufferRead dominate during heavy load Spills to disk increase latency because disk I/O is far slower than memory access Analogy: Think of ReorderBuffer as a warehouse staging floor: Small shipments move quickly in memory A huge shipment forces workers to move boxes offsite (spill-to-disk), then bring them back later, slowing everything down Why Multiple Slots Amplify the Impact The WAL is shared by all slots Each slot decodes the entire WAL stream because filtering happens after decoding Result: A single large transaction affects every slot, multiplying replication lag Recommendation: Minimize the number of slots/connectors Remember: logical_decoding_work_mem applies per slot, not globally Impact Snapshot: Scenario Spill Size I/O Impact 1 Slot 1 GB 1× I/O 5 Slots 1 GB × 5 5× I/O Lifecycle: WAL → ReorderBuffer → Memory → Spill to Disk → Read Back → Send to Client How to Detect Spills and Lag Detection should be quick and repeatable. Start by confirming slot activity and LSN distance (how far producers are ahead of consumers), then check walsender wait events to see if decoding is stalling, and finally inspect per-slot spill metrics to quantify memory overflow to disk. 1. Active slots and lag Use this to measure how far each logical slot is behind the current WAL. A large lsn_distance indicates backlog. If restart_lsn is far behind, the server must retain more WAL on disk, increasing storage pressure. SELECT slot_name, active_pid, confirmed_flush_lsn, restart_lsn, pg_current_wal_lsn(), pg_size_pretty((pg_current_wal_lsn() - confirmed_flush_lsn)) AS lsn_distance FROM pg_replication_slots; Interpretation: Focus on slots with the largest lsn_distance . If active_pid is NULL, the slot isn’t currently consuming; investigate connector health or connectivity. 2. Wait events for walsender Check whether the WAL sender backends are stalled on decoding or I/O. ReorderBuffer-related waits typically point to spill-to-disk conditions or slow downstream consumption. SELECT pid, backend_type, application_name, wait_event FROM pg_stat_activity WHERE backend_type = 'walsender' ORDER BY backend_start; Interpretation: Frequent ReorderBufferWrite / ReorderBufferRead suggests large transactions are spilling. 3. Spill stats Quantify how often and how much each slot spills from memory to disk. Rising spill_bytes and spill_count during heavy DML are strong signals to increase logical_decoding_work_mem , reduce transaction size, or tune connector throughput. SELECT slot_name, spill_txns, spill_count, pg_size_pretty(spill_bytes) AS spill_bytes, total_txns, pg_size_pretty(total_bytes) AS total_bytes, stats_reset FROM pg_stat_replication_slots; Interpretation: Compare spill_bytes across slots; if many slots spill simultaneously, aggregate I/O multiplies. Consider reducing the number of active slots or batching large DML. Fixing the Lag: Practical Strategies Once you’ve identified replication lag and its root causes, the next step is mitigation. Solutions span across the database configuration, Azure infrastructure, and the Debezium connector layer. These strategies aim to reduce I/O overhead, optimize memory usage, and ensure smooth data flow under heavy workloads. Database & Azure Layer At the database and infrastructure level, focus on reducing unnecessary overhead and ensuring resources are scaled for peak demand. Here’s what you can do: Avoid REPLICA IDENTITY FULL : prefer PRIMARY KEY; or add a unique index and set REPLICA IDENTITY USING INDEX Use appropriately scaled IO-capable storage / right SKU for higher IOPS Right-size logical_decoding_work_mem considering multiple slots Break up large DML: batch updates/deletes (10k–50k rows/commit) Schedule/throttle maintenance: stagger VACUUM/REINDEX/DDL Network placement: use Private Endpoint and co-locate Debezium/Kafka within the same region/VNet Debezium Connector Layer Connector-level tuning ensures that Debezium can keep pace with PostgreSQL WAL generation and Kafka throughput. Key adjustments include: Tune throughput & buffering: increase max.batch.size , max.queue.size , reduce poll.interval.ms Offset flush tuning: reduce offset.flush.interval.ms Heartbeats: introduce heartbeat events to detect staleness and prevent WAL buildup Conclusion Managing replication lag in Azure Database for PostgreSQL with Debezium isn’t just about tweaking parameters; it’s about understanding logical decoding internals, anticipating workload patterns, and applying proactive strategies across the entire solution. Key Takeaways: Monitor early, act fast: Use diagnostic queries to track lag, wait events, and spill activity Minimize complexity: Fewer replication slots and well-tuned connectors reduce redundant work Plan for scale: Batch large DML operations, right-size memory settings Leverage Azure capabilities: Optimize IOPS tiers, network placement for predictable performance By combining these best practices with continuous monitoring and operational discipline, you can keep your CDC pipelines healthy, even under heavy load, while ensuring downstream systems stay in sync with minimal latency. Further Reading Azure Database for PostgreSQL Flexible Server Overview PostgreSQL Logical Replication Debezium PostgreSQL Connector
Impact of CDC on SQL Server
Hi, We are testing performance impact of CDC on SQL Server. There are two identical databases (KST_S001, KST_002) on SQL Server 2017 which is running in linux container. They both have CDC enabled for 180 tables and a data generator that is doing mostly updates on these tables. The data generators are doing around 300k DML operations per minute on each database (600k total). The configuration of CDC jobs is kept default (500 transactions, 5s polling interval for capture job, I think three days retention period on cleaning job) The host machine is Azure Standard_E8-4as_v4: 4 vCPU, 64 GB RAM, 12800 IOPS, 128GB SSD. After 1 hour of running the setup with CDC enabled, the average time of 1000 updates is 73ms. With CDC disabled the average time of 1000 updates after 1 hour is 15ms. I've attached two grafana screenshots showing the difference in metrics when the CDC is disabled versus enabled on these databases. We are trying to understand the underlying mechanism that is contributing to the impact visible from the metrics. Why is memory consumption so much higher? Why do DML operations take longer with CDC enabled? CDC disabled: CDC enabled: Thanks for answers and suggestions. PeterLesson Learned #352: Could not update the metadata that indicates database enabling CDC.
Today, we got a error message while trying to enable cdc for a database using the sqladmin user. Our customer got the error message: Msg 22830, Level 16, State 1, Procedure sys.sp_cdc_enable_db_internal, Line 283 [Batch Start Line 0] Could not update the metadata that indicates database XYZ is enabled for Change Data Capture. The failure occurred when executing the command 'SetCDCTracked(Value = 1)'. The error returned was 33171: 'Only active directory users can impersonate other active directory users.'. Use the action and error to determine the cause of the failure and resubmit the request.Capture Changed Data From your Cosmos DB analytical store (Preview)
Making it super-easy to create efficient and fast ETL processing the cloud, Azure Data Factory has invested heavily in change data capture features. Today, we are super-excited to announce that Azure Cosmos DB analytics store now supports Change Data Capture (CDC), for Azure Cosmos DB API for NoSQL, and Azure Cosmos DB API for Mongo DB in public preview!
