DEV Community: Mafiree

PostgreSQL Replication Setup: Streaming Replication and Failover

Mafiree — Thu, 01 Oct 2026 11:41:22 +0000

Overview

Keeping PostgreSQL highly available and its data protected matters for modern applications. The blog presents streaming replication with failover as one of the most effective ways to achieve this, and walks through building such a system from initial setup to handling failover.

Understanding PostgreSQL Replication

Among the options available for a PostgreSQL replication setup, streaming replication is the most widely used for high availability. A standby server receives changes from the primary in real time, which keeps data consistent and makes failover possible. It relies on the Write-Ahead Log (WAL), streaming transaction logs from the primary to the replica so that every committed transaction reaches the standby with minimal delay.

Why Streaming Replication?

Real-time data synchronization
Minimal lag between primary and replica
Automatic failover through tools like Patroni
Read scaling, by offloading queries to replicas

Prerequisites

You need two PostgreSQL servers (a primary and a replica), network connectivity between them, the same PostgreSQL version on both, and SSH access to both machines.

Step-by-Step PostgreSQL Replication Setup

Step 1: Configure the primary

In postgresql.conf, set listen_addresses='*', wal_level=replica, max_wal_senders=10, max_replication_slots=10, wal_keep_size=2GB, archive_mode=on and hot_standby=on. These enable WAL streaming and let standbys connect. Next, add a line to pg_hba.conf permitting the replicator user to make replication connections from the standby's IP using scram-sha-256. Then create a dedicated replication role with the REPLICATION and LOGIN privileges and a strong password, and restart PostgreSQL to apply the changes.

Step 2: Create a replication slot

Run pg_create_physical_replication_slot('standby1'). Slots stop WAL files from being removed before the standby has received them. They are optional but strongly recommended in production.

Step 3: Initialize the standby

Stop PostgreSQL on the standby and delete its existing data directory. Clone the primary with pg_basebackup, pointing it at the primary's host, using the replicator user, and setting the target directory to $PGDATA. The command uses the flags -P, -R, -X stream, -C and -S standby1. The -R flag automatically generates the required replication configuration.

Step 4: Start the standby

Start PostgreSQL on the standby, then run SELECT pg_is_in_recovery();. A result of true confirms it is running as a standby.

Step 5: Verify streaming

On the primary, query the pg_stat_replication view for application_name, client_addr, state and sync_state. A healthy setup shows a state of "streaming" and a sync state of "async" (or "sync" if synchronous replication is configured).

Step 6: Test replication

Create a table named replication_test on the primary and insert a row. If you can see that row when querying the standby, replication is working.

Step 7: Monitor health

Regular monitoring is essential. Use pg_stat_replication and measure lag with SELECT now() - pg_last_xact_replay_timestamp();. These queries reveal delays, disconnected replicas and sync problems before they affect production. Lag should normally stay low. Persistent or growing lag can point to network bottlenecks, insufficient resources or long-running transactions on the primary.

Step 8: Failover for High Availability

Failover lets a standby be promoted to primary if the primary becomes unavailable, which limits downtime and keeps the service running.

Best Practices for PostgreSQL Failover

Use high-availability tools such as Patroni to automate failover and manage the cluster.
Test failover and failback procedures regularly.
Keep regular backups of both primary and standby servers.
Continuously monitor replication lag, server health and connectivity, with alerts for anomalies.
Document failover and recovery procedures so you can respond quickly during outages.

Manual failover Without automatic failover, you can promote the standby by running SELECT pg_promote();. Afterward, SELECT pg_is_in_recovery(); should return false, confirming the server is now acting as the primary.

Mafiree's PostgreSQL Replication Services

Mafiree offers PostgreSQL DBA services covering replication setup, performance monitoring and smooth failover. Its experts can help you design and implement replication strategies, monitor replication health and performance, implement automated failover, and ensure compliance with data protection standards. The blog also points readers to a separate Mafiree guide on choosing between streaming, logical and other replication types.

Conclusion

A dependable replication system is crucial for data integrity and business continuity. Streaming replication gives near real-time synchronization between primary and standby servers, and it supports high availability, disaster recovery and read scalability. Proper configuration, monitoring and failover procedures are what maximize uptime and reduce risk. Mafiree positions itself as able to help with new deployments or improvements to existing systems.

PostgreSQL 19 REPACK: Removing Table Bloat Online

Mafiree — Thu, 24 Sep 2026 11:44:46 +0000

PostgreSQL 19 REPACK is a native command that unifies VACUUM FULL and CLUSTER and adds a CONCURRENTLY option, so you no longer have to choose between reclaiming disk space and keeping a table online. After a large delete or an update-heavy workload, a table can bloat, and autovacuum can't return that space to the operating system. DBAs have had two poor choices: VACUUM FULL, which locks the table for the whole rewrite, or the third-party pg_repack extension, with the hope that version, permissions and maintenance window all line up.

Why the Old Tools Fell Short

VACUUM FULL rewrites a table into a new file, discards dead tuples and returns the space to the OS. CLUSTER does the same but also reorders rows by an index, which helps range scans. Both take an ACCESS EXCLUSIVE lock for the entire rewrite. On a table of several hundred GB, that is an outage window that must be scheduled and approved, with no guarantee it finishes on time. This is why pg_repack became a standard tool: it builds a shadow table, captures changes with triggers, and swaps the tables at the end.

What REPACK Is

Both older commands were always doing the same job: copying live tuples into a new file and swapping it in. They differed only in whether index order was preserved. REPACK treats them as two modes of one operation, which gives the project a single place to add features like concurrency. It copies live rows, rebuilds every index into new files, and swaps the new files in.
The blog gives the syntax as of Beta 2. The options are VERBOSE, ANALYZE and CONCURRENTLY. Plain REPACK table; behaves like VACUUM FULL, adding USING INDEX idx makes it behave like CLUSTER, and specific columns can be named, for example to repack storage after dropping a wide column. REPACK (VERBOSE) shows progress output, and REPACK (ANALYZE) runs ANALYZE after the rewrite. The \h repack command in psql shows the full syntax, and the official PostgreSQL 19 release notes and REPACK documentation are the sources of record as the beta evolves.

How CONCURRENTLY Works

With CONCURRENTLY, REPACK does not hold an ACCESS EXCLUSIVE lock while it copies the table. Any inserts, updates and deletes that happen during the copy are captured through logical decoding and replayed onto the new copy just before the final file swap. An exclusive lock is needed only for that last step, and it is held briefly.

The Hands-On Test

The author built a one-million-row table with a primary key and two secondary indexes, then simulated churn: every row updated twice, then half, then a third, followed by a 10% delete. The table used 1341 MB, its indexes 241 MB, and the total was 1582 MB, despite holding under a million live rows.
REPACK (VERBOSE, ANALYZE) took about 27 seconds and reported removable and nonremovable row versions, plus CPU, buffer and WAL statistics. Afterward the table was 370 MB, the indexes 89 MB, and the total 459 MB.

Monitoring with pg_stat_progress_repack

This new view lets you watch a running repack from another session. It exposes the phase, command, relation ID, index in use, heap tuples scanned, inserted, updated and deleted, block progress, and index rebuild count. The phase column tells you which stage the job is in, so you don't have to guess whether it's stuck.

Limitations

REPACK (CONCURRENTLY) cannot be used when:
the table is UNLOGGED;
the table is partitioned (individual partitions can still be repacked one at a time);
the table lacks a primary key and index-based replica identity;
the command runs inside a transaction block.
Any repack, concurrent or not, requires the MAINTAIN privilege on the table. This matters if you run it as a lower-privileged migration or automation user, and superusers need no extra permission changes.

Should You Keep pg_repack?

Yes. pg_repack is the only online option on PostgreSQL 18 and earlier. On PostgreSQL 19, its per-partition batch tooling and its track record on very large, long-running jobs still make it worth keeping for edge cases the new core command doesn't cover yet, particularly partitioned tables as a whole.
For the common case of a single logged table with a primary key bloated by update or delete churn, native REPACK (CONCURRENTLY) removes a dependency and a maintenance window, and gives you a progress view for free.

Conclusion

Bloat has always been well understood. What was missing was a fix that didn't force a trade-off between reclaiming space and keeping the table online. VACUUM FULL gave the first at the cost of the second, and pg_repack gave both but required an extension installed and version-matched on every environment. REPACK in PostgreSQL 19 removes that trade-off inside the core, with a progress view so you aren't flying blind.

VPN Troubleshooting Case Study: Tunnel UP, Monitoring Failing

Mafiree — Fri, 18 Sep 2026 06:29:21 +0000

Intermittent VPN problems are notoriously difficult to diagnose during VPN troubleshooting because everything can look fine on the surface while something is quietly going wrong underneath. Mafiree encountered exactly this scenario with a client whose infrastructure was monitored through a jump server connected via a site-to-site IPsec VPN. The tunnel showed up, routing worked correctly, SSH access was available, and packet loss was negligible - yet the monitoring dashboards kept showing gaps and interruptions. The team was called in to find out why, initially assuming the fault lay outside the VPN itself. That assumption proved incorrect.

VPN Troubleshooting: Why Tunnel Status Can Be Misleading

The investigation's central lesson was that an "established" VPN connection can still hide serious problems. A tunnel can remain up while suffering from rekeying failures, fragmentation issues, mismatched lifetime settings, failed Phase 2 negotiations, or firmware bugs - all of which can cause brief disruptions that are too short to register as a full outage but long enough to break monitoring data collection.

Checking MTU and Encryption Overhead

Because the problem emerged after VPN-related changes, the team first suspected MTU issues. Testing showed that 1472-byte packets failed while 1400-byte packets went through, confirming that encapsulation overhead was cutting into the usable payload. Given the IPsec setup - AES-256 encryption, SHA-256 authentication, and IKEv2 with Perfect Forward Secrecy - the effective payload size was roughly 1365 bytes. This was a real factor, but it didn't account for the intermittent nature of the failures, since smaller packets and control traffic kept working normally.

Network Path Analysis Using MTR

To check whether unstable routing was involved, engineers ran an extensive 100-packet MTR test. The results were clean across the board: zero packet loss, stable latency with no jitter, consistent round-trip times, and no route changes. The network path itself was healthy, so attention shifted elsewhere.

SSH Tests Looked Fine, At First

Since monitoring depended on the jump server, SSH performance was tested repeatedly. Ten separate sessions connected successfully in under a second each time, suggesting authentication and tunnel responsiveness were both working. But this created a contradiction: the tunnel tested as healthy, yet the actual service relying on it kept failing in production.

Finding the Hidden Pattern

Rather than continuing with occasional manual checks, the team built a continuous SSH testing script with detailed logging. This revealed something spot-checks had missed entirely: the failures weren't random. They occurred at recurring intervals, hinting at a connection to tunnel lifetime settings or Security Association renegotiation cycles - a strong clue that timing itself was significant.

Tracing It Back to a Firmware Change

Reviewing the change history showed the problem started right after the client upgraded their firewall firmware and adjusted VPN settings to match the new recommendations. This redirected the investigation toward the VPN configuration itself. Every parameter - Phase 1 and Phase 2 encryption, preshared keys, PFS settings, and lifetime values - was checked and found to match on both ends. Even after further adjustments, though, the intermittent issues continued, showing that matching configurations alone weren't the whole story.

The Real Cause: Silent Configuration Drift

The breakthrough came from examining firewall syslogs instead of the VPN status dashboard. The logs showed that VPN parameters were being automatically regenerated and reverted to firmware defaults during operation - even after administrators had manually corrected them - with no indication of this happening on the dashboard. This explained why the tunnel stayed "up," connectivity looked mostly normal, and yet monitoring kept breaking: the real issue was a firmware-driven configuration consistency problem invisible to standard status checks.

VPN Troubleshooting Resolution

As a short-term fix, Mafiree aligned tunnel settings with the values the firmware kept reverting to, which improved but didn't fully resolve things. The client eventually replaced the firewall entirely, after which monitoring stabilized immediately, VPN interruptions stopped, and no further anomalies appeared.

Key Takeaways

The case reinforced several lessons: an "UP" status doesn't guarantee health; recent changes should always be checked first; continuous testing catches patterns that spot checks miss; logs often reveal what dashboards hide; and rigorous, evidence-based troubleshooting rather than assumptions - is what ultimately uncovers root causes.

MySQL Deadlock Analysis: Diagnosing and Resolving Lock Contention in High-Concurrency Workloads

Mafiree — Thu, 10 Sep 2026 05:29:13 +0000

What Is a MySQL Deadlock?

A MySQL deadlock happens when two or more transactions each hold a lock that another transaction needs, forming a circular dependency that stops all of them from moving forward. InnoDB automatically spots these cycles and rolls back one of the transactions, sending back error 1213 to the affected application, which should be built to retry safely. This isn't a bug - it's an expected outcome of concurrent writes competing for the same data, and it becomes a real concern only when it happens too often under heavy load.

How Do You Diagnose a MySQL Deadlock?

Diagnosis starts with running SHOW ENGINE INNODB STATUS\G and examining the "LATEST DETECTED DEADLOCK" section, which shows the transactions involved, the locks they held and wanted, and the SQL that triggered the conflict. For ongoing monitoring, MySQL exposes live lock data through system tables, but the right ones depend on version: MySQL 8.0+ uses performance_schema.data_locks, performance_schema.data_lock_waits, and information_schema.INNODB_TRX, while MySQL 5.7 and earlier relies on the now-removed INFORMATION_SCHEMA.INNODB_LOCKS and INNODB_LOCK_WAITS. Turning on the slow query log (with long_query_time = 0 and log_slow_admin_statements) also helps catch long-running queries that contribute to lock contention.

Why Do MySQL Deadlocks Happen?

Common causes include: transactions reaching the same rows through different indexes (creating inconsistent lock order), transactions that stay open too long and hold locks longer than necessary, and applications that access tables in inconsistent orders across different code paths.

How Do You Fix a MySQL Deadlock?

InnoDB's deadlock detector watches for cycles in the wait-for graph. When one appears, it picks a "victim" transaction to roll back, weighing it by how many rows it has modified and how much work it has already done — generally sparing the transaction with more invested work. This is described as a heuristic rather than a guarantee, so applications shouldn't assume they can predict which transaction will be chosen.

How Should an Application Handle MySQL Error 1213?

Once a transaction is rolled back, the app receives error 1213. This should be treated as a normal, recoverable event, not a failure — the fix is to reissue the entire transaction using retry logic with exponential backoff.

How Can You Prevent Recurring MySQL Deadlocks?

Prevention comes down to three practices: keep queries consistent in which indexes they use (avoiding full scans), keep transactions short by committing early and often, and always access tables in the same order across all transactions to avoid circular waits.

MySQL Deadlock Diagnostic Workflow

A nine-step workflow is laid out: capture the deadlock evidence immediately (since it gets overwritten), identify the two transactions and note which was the victim, map out each transaction's held versus requested locks, trace those locks back to the actual SQL and indexes via EXPLAIN, confirm there's a genuine circular dependency (otherwise it may just be a lock wait timeout, error 1205), apply the appropriate fix, test the fix under real concurrent load, keep monitoring with innodb_print_all_deadlocks enabled, and maintain retry logic as an ongoing safety net.

MySQL Deadlock Troubleshooting Example

A quick-reference table matches symptoms to causes: an isolated error 1213 usually points to inconsistent index usage; deadlocks clustering during peak load usually mean long transactions under concurrency; and deadlocks recurring across varied queries usually mean inconsistent lock ordering.

Mafiree Verdict: What Should You Fix First?

The recommended fix order is: enforce consistent locking order first, then shorten long transactions, then align index usage - with retry logic always present as a backstop, not a primary solution. This kind of structured diagnostic discipline is also central to how Mafiree's Managed Database Services team approaches recurring lock contention issues for clients.

Case Study: Resolving High-Concurrency Deadlocks at Scale

A case study describes an e-commerce client whose deadlocks stemmed from inconsistent index usage during peak hours. Mafiree standardized locking order, broke slow transactions into smaller pieces, and added monitoring/alerting, resulting in roughly a 70% drop in deadlock occurrences (results specific to that engagement, varying by workload).

What Are the Limitations of Deadlock Prevention?

Deadlocks can never be fully eliminated - even well-tuned systems will occasionally hit one so the real goal is making them rare and predictable, with safe retries covering what slips through.

Conclusion

Deadlock analysis is essential for stable, high-performance MySQL systems. Understanding why deadlocks happen and using the right diagnostic tools lets teams address issues proactively, whether running a small app or a large enterprise system.

TiDB Out of Memory (OOM) Errors: Root Cause Analysis & Memory Tuning Guide

Mafiree — Fri, 04 Sep 2026 07:07:16 +0000

Overview

TiDB Out of Memory errors have two distinct forms, and figuring out which one occurred is the essential first step in troubleshooting. The immediate fix is usually a simple restart, but understanding the actual cause is what prevents recurrence.

There Are Two Different TiDB OOM Situations

In the first case, the Linux operating system runs out of available memory and kills the tidb-server process, dropping active connections and terminating running queries before TiDB restarts - this is the more serious scenario since it affects the entire server. In the second case, TiDB's own internal memory controller cancels a single runaway query (returning an "Out of Memory Quota" error) while the node itself stays healthy. This is the protection system functioning as intended, though the underlying cause still deserves investigation.

How TiDB Memory Controls Work

TiDB's memory thresholds operate in stages: 0–70% is the normal operating range; 70–80% is a good point for early-warning alerts by setting tidb_memory_usage_alarm_ratio to 0.7; and 80%+ is the default tidb_server_memory_limit threshold where TiDB begins killing queries.

First: Confirm What Actually Happened

Before changing any configuration, determine which type of OOM occurred. Check OS logs with dmesg -T | grep tidb-server for oom-killer activity, look at the "Welcome to TiDB" restart line in tidb.log as a timestamp anchor, and cross-check the Grafana memory usage graph under TiDB → Server → Memory Usage - a pattern that climbs steadily, drops to zero, then climbs again typically signals a process restart.

What Usually Causes TiDB OOM Errors?

The most common root causes are memory-heavy queries with large intermediate result sets (the most frequent cause), too many concurrent sessions collectively exhausting available RAM (common), memory not being released properly over time (occasional), and under-provisioned hardware (less common). The first two account for most production incidents.

SQL Patterns That Consume Large Amounts of Memory

HashJoin against a large table: builds a huge in-memory hash table on the inner side and can consume gigabytes. Checking EXPLAIN for large estRows on the inner side and hinting toward MergeJoin can reduce memory use, since MergeJoin works with sorted streams instead of a hash table.
HashAgg on large grouped datasets: builds per-worker hash tables for grouped data, which can be replaced with StreamAgg for lower memory use by processing sorted rows.
Stale statistics causing bad query plans: outdated statistics can cause the optimizer to badly misjudge row counts (e.g., estimating 1,000 rows for a 10-million-row scan) and pick memory-hungry algorithms. Checking SHOW STATS_HEALTHY and running ANALYZE TABLE regularly on high-write tables helps.
Large transactions holding too much memory: TiDB caches all writes in memory before commit, so a transaction touching millions of rows can use two to three times the actual data size. Breaking bulk deletes/updates into smaller batches (with a brief 50–100ms sleep between rounds) or using tidb_dml_type = "bulk" or non-transactional DML mitigates this.

TiDB Memory Parameters Worth Configuring

TiDB memory protection operates in layers — per-query, per-instance, and OS/cgroup limits. Four parameters matter most:

tidb_mem_quota_query (default 1GB, session/global): caps memory per query.
tidb_mem_oom_action(default CANCEL, session/global): CANCEL terminates the offending query and is right for production; LOG lets the query continue while logging the event, useful only temporarily for investigation.
tidb_server_memory_limit (default 80% of system memory, v6.5.0+): caps the whole TiDB process; should be set explicitly in hybrid deployments sharing a server with other workloads.
tidb_memory_usage_alarm_ratio (default 0.8, global): triggers diagnostic collection; setting it to 0.7 gives earlier warning.

TiDB Disk Spill: Let TiDB Use Disk Instead of RAM

Disk spill lets TiDB offload intermediate execution data (from Sort, MergeJoin, HashJoin, HashAgg, or TopN operators) to disk under memory pressure rather than crashing, governed by tidb_mem_quota_query, tidb_enable_tmp_storage_on_oom, tmp-storage-path, and tmp-storage-quota. Spill support has improved across versions, especially for HashAgg, so behavior should be verified per release. A dedicated tmp-storage-path is recommended, but spill is a safety net, not a substitute for query optimization.

View TiDB Memory Usage Through INFORMATION_SCHEMA

TiDB exposes MEMORY_USAGE, CLUSTER_MEMORY_USAGE, MEMORY_USAGE_OPS_HISTORY, and CLUSTER_MEMORY_USAGE_OPS_HISTORY, with the OPS_HISTORY tables retaining the latest 50 records per instance for post-incident analysis.

Production Readiness Checklist

Memory configuration: explicit tidb_server_memory_limit, workload-based tidb_mem_quota_query, tidb_mem_oom_action set to CANCEL, appropriate alarm ratio.
Query optimization: review with EXPLAIN ANALYZE, check HashJoin/HashAgg, refresh stale statistics, monitor high-memory queries.
Transaction management: batch large DELETE/UPDATE operations, avoid unnecessary million-row transactions, evaluate bulk/non-transactional DML.
Disk spill: dedicated temp storage path, sufficient disk space, monitor spill usage.
Infrastructure: avoid undersized hardware, account for co-located services, configure cgroup boundaries in hybrid deployments.
If recurring OOM issues persist despite these steps, bringing in DBA consulting services can provide the deeper diagnostic expertise needed to resolve them.

Conclusion

TiDB OOM errors are rarely random - they trace back to specific SQL patterns, undersized memory limits, or mismatched defaults. Confirming whether the OS or TiDB's own memory manager caused the event is the first troubleshooting step; from there, tuning tidb_mem_quota_query and tidb_server_memory_limit, optimizing costly queries, and enabling disk spill as a safety net prevents most repeat incidents.

MySQL Consulting vs In-House DBA: Which Is Right for Your Team?

Mafiree — Wed, 26 Aug 2026 09:08:47 +0000

Choosing between an in-house DBA and MySQL consulting isn't about picking a universal winner - it depends on how much database work your organization needs on an ongoing basis, how specialized that work is, and how much coverage you require. Understanding MySQL Consulting vs In-House DBA trade-offs is key: in-house DBAs bring deep institutional knowledge and daily hands-on ownership, while consultants offer flexible access to specialized skills for tasks like performance tuning, migrations, and high-availability projects. Many growing companies find that a hybrid approach - internal staff handling routine work, with outside experts called in for complex needs - offers the best of both worlds.

In-House DBA Teams

An in-house DBA is embedded directly in your organization and understands your infrastructure, workflows, and business needs intimately. The benefits include familiarity with internal systems (allowing low-disruption changes), immediate availability without communication delays, and long-term stability since the team knows the environment thoroughly. The downsides are limited exposure to cutting-edge tools and methods, higher overhead from salaries, benefits, training, and equipment, and the risk of skill gaps as technology evolves faster than internal staff can keep pace.

MySQL Consulting Services

Firms like Mafiree bring specialized, cross-industry experience and scalable solutions without the cost of maintaining a full-time specialist team. Advantages include access to broad expertise gained from varied environments, cost efficiency since you only pay for time actually used, and access to current tools and best practices consultants tend to adopt early. Drawbacks include a learning curve as consultants get up to speed on your specific setup, potential communication friction from remote collaboration, and dependency risk if you rely too heavily on external relationships.

Cost Comparison

Costs break down differently for each model: in-house DBAs involve salary and benefits, recruitment and onboarding, training investment, on-call coverage, project work absorbed into salary, and separately purchased tooling. Consulting instead involves hourly, monthly, or AMC fees, no onboarding costs, skills upkeep included by the provider, emergency coverage, separate project fees, and often bundled tooling. These figures are illustrative and vary by region, seniority, and workload.

Key Decision Factors

Four factors matter most: budget and resource allocation (upfront investment vs. predictable per-project costs), technical complexity (niche expertise may exceed in-house capability), scalability needs (consulting allows flexible scaling during growth or peak periods without permanent hires), and risk tolerance (in-house suits organizations wanting control and familiarity, while consulting suits those prioritizing agility).
Recommended models by scenario: early-stage startups and migration projects or performance crises fit consulting; SaaS companies with 24/7 workloads and enterprises with existing DBA teams benefit from in-house plus consulting; stable, low-complexity workloads suit an in-house or generalist approach.

Decision Framework

A simple rule of thumb: stable workload with continuous ownership needs → in-house DBA; specialized, temporary problems → consulting; stable workload with occasional specialist needs → hybrid model; high availability with limited internal coverage → managed database services.

Mafiree's Verdict

The decision should rest on workload and capability gaps rather than a one-size-fits-all rule. Continuous, deeply integrated database operations favor an in-house DBA. Specialized expertise, extra capacity, or complex project support favor consulting. Organizations needing both can adopt a hybrid model, and those wanting ongoing coverage plus specialist expertise may consider managed database services instead of expanding their internal team.

Mafiree's Managed Database Services

Mafiree offers a hybrid approach combining in-house expertise with external consulting, including 24/7 monitoring and support, performance optimization and tuning, disaster recovery planning, migration assistance from legacy systems, and security audits with compliance management.

Conclusion

There's no universal answer - the right choice depends on budget, technical demands, growth plans, and long-term strategy. Whether building an internal team or partnering with a consulting firm, the goal is an informed choice supporting database performance and business success. Mafiree positions its managed services as a balanced option offering both specialized knowledge and consistent support without the overhead of full-time hires.

Understanding Software Deployment Environments: DEV, QA, UAT & PROD

Mafiree — Tue, 18 Aug 2026 08:19:09 +0000

Since software is updated constantly, releasing changes without properly separated environments risks breaking things for live users, leaking sensitive data, or disrupting core business functions. Software deployment environments give developers, testers, and business stakeholders controlled stages to validate an application before it reaches customers. Within DevOps and CI/CD practices, these environments help teams catch bugs earlier, strengthen security compliance, protect production systems, cut down on failed deployments, and ship updates more quickly and safely - a need shared by startups pushing weekly updates and enterprises running mission-critical systems.

Why Separate Environments Matter

Every application moves through a sequence of distinct stages before reaching end users: DEV → QA → UAT → PROD. Far from being an unnecessary process, this separation is what allows teams to move quickly without putting real users at risk.

Development (DEV)

DEV is where coding, feature-building, and initial testing happen. It's a fast-moving space where stability takes a back seat to experimentation - developers constantly add features, fix issues, and test how components integrate. To mirror application behavior without needing full infrastructure, teams typically rely on local tools like Docker, local databases, Minikube (a lightweight Kubernetes setup), virtual machines, and local API servers.

QA / Test Environment

Here, the priority shifts to confirming the software works correctly before it reaches business users. QA engineers run functional, regression, API, security, performance, and automated tests to verify new features work properly and that recent changes haven't broken existing functionality.

UAT (User Acceptance Testing) / Staging

UAT is the last checkpoint before production. This environment is built to closely mirror production - matching configurations, database structures, load balancers, security policies, and network setup - so teams can catch issues that earlier stages might miss. Unlike QA, which is technically focused, UAT centers on business validation: stakeholders, product owners, and end users check whether the application supports real workflows like order processing, registration, payments, reporting, and approvals.

Production (PROD)

This is the live environment serving actual users, so reliability, security, performance, and availability take top priority since problems here directly affect customers and business operations. Production infrastructure typically includes auto-scaling, load balancers, backups, monitoring dashboards, security tools, and incident response systems. To limit risk when rolling out changes, teams use techniques such as Blue-Green Deployments, Canary Releases, Rolling Updates, and Feature Flags. Because failures can cause revenue loss, reputational harm, security breaches, or compliance issues, production releases go through strict change management involving approvals, rollback plans, and close monitoring.

Mafiree's Deployment Flow Example

Mafiree illustrates its own process using a checkout feature: developers build and test it locally in DEV; QA engineers verify payment flows, API behavior, and security in QA; business teams confirm the checkout experience works for customers in UAT; and finally, the feature ships to live users through a controlled release in PROD. This staged approach catches problems before customers ever encounter them.

Security Practices for Software deployment environments dev qa uat & prod github

Recommended safeguards include using separate credentials per environment, never using production data in DEV, masking sensitive customer data, restricting production access, enabling logging and monitoring, applying Infrastructure as Code, and using role-based access control. Security tightens progressively as code moves closer to production.

Common Pipeline Challenges

Teams often run into configuration drift between environments, inconsistent test data, failed deployments, dependency mismatches, weak rollback planning, and slow testing cycles — issues that automation and solid DevOps practices help resolve.

Staging vs. UAT

Staging is technical in focus, used by developers and testers to catch bugs and rehearse releases through smoke tests. UAT is business-focused, used by clients, product owners, and analysts to confirm real-world scenarios work as intended. Once UAT is signed off, the release moves to production.

Environment Comparison (DEV vs SIT vs UAT vs PROD)

The stages differ across stability (low to highest), data type (synthetic to live), users (developers to end users), deployment frequency (every commit to controlled), risk (high to zero), monitoring (minimal to 24/7), and configuration (flexible to strict).

Conclusion

Separating development, testing, validation, and production environments helps organizations lower deployment risk, boost software quality, safeguard customer data, speed up releases, and maintain business continuity - making well-managed environments a core foundation of modern software engineering as DevOps and cloud-native practices continue to evolve.

ProxySQL vs HAProxy for MySQL High Availability

Mafiree — Thu, 13 Aug 2026 10:54:46 +0000

When building reliable database infrastructure, ProxySQL vs HAProxy is a common decision point for teams managing MySQL traffic. Both tools help improve database reliability through load balancing, failover support, and query routing, but they take different approaches.

What Is ProxySQL

ProxySQL is an open-source, high-performance proxy built specifically for MySQL and MySQL-compatible databases. It sits between applications and database servers, offering read/write splitting, query caching, load balancing, failover handling, and performance monitoring.

What Is HAProxy

HAProxy is a widely used open-source load balancer that works across HTTP, HTTPS, TCP, and UDP protocols. Though originally built for web traffic, it has since been adapted to handle database connections too. Its core strengths include multi-protocol support, advanced load-balancing algorithms, health checks with failover, and SSL termination plus compression.

Key Differences

Protocol support: ProxySQL is purpose-built for MySQL, with deep integration features like query rewriting, transaction handling, and connection pooling. HAProxy covers a wider protocol range (HTTP, HTTPS, TCP, UDP) but doesn't offer the same MySQL-specific depth.
Load balancing: ProxySQL routes queries intelligently based on server status and query type. HAProxy relies on solid TCP-level load balancing using algorithms such as round-robin and least connections.
Read/write splitting: This is where ProxySQL stands out - it natively routes queries to the right server (primary or read replica) and supports configuration changes without downtime. HAProxy generally needs extra configuration or external tools to achieve the same result, making it less streamlined here.
Query caching and optimization: ProxySQL includes built-in caching for SELECT query results, query rewriting for performance gains, and detailed metrics. HAProxy has no built-in caching or optimization layer, since it operates at the transport level.
Failover and high availability: Both support failover, but differently. ProxySQL bases failover decisions on server health and query execution status, giving it database-aware sophistication. HAProxy uses health checks to redirect traffic to backup servers but lacks that same database-level awareness.
Performance monitoring: ProxySQL offers detailed, MySQL-specific monitoring — query execution stats, connection tracking, and server health reports. HAProxy's metrics are more generic and transport-focused rather than database-specific.

When to Choose Each Tool

Choose ProxySQL if MySQL-specific performance tuning is a priority, read/write splitting is essential, you need advanced query caching/rewriting, or your setup depends heavily on MySQL replication topologies.
Choose HAProxy if you're managing traffic across multiple protocols (HTTP, HTTPS, TCP), need general-purpose load balancing across several services, require SSL termination and compression, or prefer a lightweight, generic proxy.

On Performance Claims

The blog notes that while ProxySQL is designed to boost MySQL performance through intelligent routing, connection pooling, caching, and read/write splitting, actual gains vary by workload — so it recommends workload-specific benchmarking rather than relying on universal performance numbers.

Comparison Summary

Protocol Support
ProxySQL: MySQL-specific
HAProxy: TCP/UDP, HTTP/HTTPS
Read/Write Splitting
ProxySQL: Native
HAProxy: Requires external tools
Query Caching
ProxySQL: Built-in
HAProxy: None
Failover Handling
ProxySQL: Advanced, database-aware
HAProxy: Basic health checks
Performance Monitoring
ProxySQL: Detailed MySQL metrics
HAProxy: Generic transport stats

Bottom Line

The choice comes down to your infrastructure needs. If MySQL-specific features like read/write splitting, caching, and performance tuning matter most, ProxySQL is the stronger fit. If you need a general-purpose, multi-protocol load balancer for managing traffic across various services, HAProxy offers more flexibility.

MySQL Consultant: When to Call in a Performance Expert

Mafiree — Thu, 06 Aug 2026 04:12:07 +0000

Most MySQL slowdowns can be handled by an internal team - a missing index here, a bloated table there. But a smaller set of problems don't respond to routine tuning: they keep coming back, spread across systems, or carry risks a team hasn't dealt with before. Knowing where that line sits helps avoid letting an incident force the decision.

Signs It's Beyond Normal Troubleshooting

Five symptoms are common enough that most teams encounter at least one. Individually they're often manageable, but they become escalation signals when persistent, recurring, or stacking together:

1. Query times keep climbing — queries that once ran in milliseconds now take seconds, and routine fixes don't reverse the trend.

2. CPU or memory stays maxed — resource usage sits near capacity persistently, not just during expected traffic spikes.

3. Lock waits keep recurring — transactions repeatedly queue behind each other even after indexes and query shapes have been reviewed.

4. User-facing response times slip — the slowdown becomes visible to customers or internal users, not just in the database layer.

5. Replication lag keeps growing — primary-to-replica lag trends upward with no obvious single cause.
The key warning sign: if two or more of these show up at once, if a fix works briefly then regresses, or if no one can explain why the last change helped, the root cause likely runs deeper than a single query or config setting.

In-House Fix vs. Calling a Consultant

The article offers a quick reference: minor query slowdowns and index reviews are generally manageable in-house. Repeated lock waits, replication lag, server tuning across varied workloads, and migration-related performance risk are situations where bringing in a consultant is recommended.

What a Consultant Checks First

A good MySQL consultant starts with a system-wide review rather than jumping straight to the slowest query, to determine whether that query is the actual cause or just a symptom. This includes:

Slow query logs and execution plans, to separate outliers from patterns
InnoDB buffer pool sizing and hit rate versus actual working-set size
Lock and transaction behavior under real concurrency, not synthetic load
Replication topology and applier configuration, where relevant
Server parameters checked against the actual workload, not generic defaults

When to Call One Immediately

Certain project types justify bringing in a specialist from the outset rather than after a failure:
Complex query optimization — large datasets, multi-way joins, subqueries, or aggregations often hide optimization paths not obvious from a single EXPLAIN.
Server configuration issues — buffer sizes, connection limits, and log settings interact, so tuning one in isolation often just shifts the bottleneck elsewhere.
Scalability challenges — when data volume or concurrent users outpace capacity, the fix is usually architectural rather than a config tweak.
Migration or upgrade projects — version upgrades and architecture changes risk regressions that only appear under production load.
Security and compliance needs — access controls, encryption, and monitoring involve performance trade-offs that are easy to misjudge without experience.
Some quick figures from the piece: unresolved issues often cost teams 2–4 weeks silently before escalation; two or more stacked signals is typically the threshold for bringing in a specialist; and most diagnostic and tuning work requires zero downtime.

What the Business Gains by Acting Early

Bringing in a consultant before an incident — rather than during one — changes the engagement. Teams that act early typically see: a diagnosis before the problem becomes customer-visible, fewer total engineering hours than repeated internal troubleshooting, a scaling plan rather than a patch that resurfaces later, and greater confidence heading into migrations, launches, or growth events.

How Mafiree Helps

Mafiree's database consultants run full audits before recommending fixes, covering performance audits and diagnostics, query optimization and execution plan analysis, server configuration tuning, index optimization strategies, database architecture reviews, and migration planning and implementation support.

Conclusion

Most MySQL issues are within reach of a capable internal team. The exceptions share a pattern - they recur after being "fixed," touch more than one system, or carry more risk than the team has handled before. Spotting that pattern early, rather than after an outage, is what keeps a performance issue from turning into a business problem.

Building a Real-World AWS DevOps Agent: From Concept to Implementation

Mafiree — Wed, 29 Jul 2026 11:39:09 +0000

An AWS DevOps Agent is an intelligent automation layer built on top of AWS that continuously monitors, analyzes, and manages cloud infrastructure and application workflows. Rather than relying on manual intervention or static scripts, it functions like a virtual DevOps engineer - handling deployments, detecting issues, and responding to events in real time. Using cloud-native services and event-driven architecture, it helps teams move faster while keeping systems stable and reliable. Organizations often pair this kind of automation with broader DevOps services and managed database consulting to build reliable, scalable cloud environments.

What the AWS DevOps Agent Can Do

The agent works as an intelligent observer and analysis layer over AWS. Although it integrates with services like AWS CodePipeline, CodeBuild, and CodeDeploy, its core strength isn't executing pipelines directly - it's analyzing, monitoring, and improving them. It continuously watches CI/CD workflows, helping teams investigate failures, understand deployment issues, and pinpoint what changed between successful and failed runs. Rather than manually combing through logs, engineers can let the agent analyze pipeline history, surface errors, and highlight likely root causes, making troubleshooting faster in complex environments.

Investigation and Monitoring

The agent plays a central role in monitoring and troubleshooting production systems. Using telemetry from Amazon CloudWatch, it analyzes infrastructure metrics, logs, and application behavior in real time, detecting anomalies like CPU spikes, rising latency, or unusual error rates. Beyond basic monitoring, it can correlate logs across systems, analyze trends, and trace requests through distributed architectures, providing actionable insights that cut down the time needed for root cause analysis during deployment failures, performance issues, or incidents. This works best alongside comprehensive monitoring services offering proactive observability across applications, infrastructure, and databases.

Turning Alerts into Root Cause Analysis

Traditional monitoring tools often produce alerts without context, leaving engineers to investigate manually. The agent closes this gap by converting raw alerts into meaningful insights - correlating metrics, logs, and events to determine whether an issue stems from infrastructure, application code, or a deployment change. It can also integrate with multiple observability tools, combining data from CloudWatch and third-party platforms like New Relic, giving teams a unified view instead of forcing them to switch between tools during an incident. It further automates parts of the investigation itself, checking logs, metrics, and recent deployments automatically to flag anomalies and possible causes. For instance, if a web application slows down, the agent can determine whether the cause is high resource use, a recent deployment, or a dependent service issue.

Chat-Based Interaction

A standout feature is the agent's chat interface, letting engineers ask natural-language questions like "Why did the deployment fail?" or "What caused the CPU spike?" instead of digging through dashboards and logs manually. Built on AWS services, this conversational layer understands context and pulls relevant data from across the environment. It also supports guided investigation, suggesting follow-up questions and helping engineers drill down into issues step by step, making incident response faster and reducing the need for deep manual debugging.

What the AWS DevOps Agent Does Not Do

Importantly, the agent does not directly execute changes in the environment. Its ability to act is bound by specific architectural and security limits, making it a decision-support system rather than a fully autonomous executor. It delivers insights, recommendations, and reports so engineers can act quickly and confidently themselves.

Real-World Example

Consider an ecommerce application on AWS deployed via CodePipeline, hosted on Amazon ECS, and monitored through CloudWatch. A developer pushes code, CodePipeline builds and deploys it, and CloudWatch continuously gathers metrics and logs. During peak traffic, CloudWatch detects a CPU spike and slower response times. The DevOps Agent receives this event and begins an automated investigation - analyzing recent deployments, reviewing logs, comparing metrics against historical trends, and correlating infrastructure events. It determines the slowdown began right after the latest deployment and traces it to a newly introduced database query. When an engineer asks via chat why the app is slow, the agent summarizes its findings, identifies affected services, and recommends reviewing the database changes or rolling back the deployment.

TiDB Pause Resume DDL: Safely Manage Long-Running Schema Changes in Production

Mafiree — Wed, 08 Jul 2026 05:40:26 +0000

Running schema changes on production databases is risky, especially on large, high-traffic tables. A common scenario: an ALTER TABLE command to add an index seems to run fine, but as it nears completion, query latency spikes and users notice slower performance. Canceling a long-running DDL job at that point isn't ideal, since it wastes all progress made and forces a restart from scratch. TiDB's pause and resume DDL feature solves this by letting administrators pause a job during peak load and resume it later, minimizing impact on production.

What is TiDB DDL Pause and Resume?

TiDB allows running DDL jobs to be paused and resumed without cancellation, which is useful when workloads fluctuate or during emergencies. In a TiDB cluster, one node acts as the "DDL owner," coordinating the workers that execute schema changes. The current owner can be identified with ADMIN SHOW DDL;, which returns the schema version, owner ID, and owner address.

How DDL Blocking Impacts Production Queries

Even though TiDB supports online DDL, operations like index creation or column-type changes still consume real resources while running. These "physical DDL operations" (which modify both metadata and underlying data) can cause increased write amplification, higher CPU usage, extra storage I/O, and higher query latency during the reorganization phase — commonly seen when creating indexes on large tables or backfilling large datasets.

Pausing a DDL Job — Step by Step

Create the DDL job, e.g., CREATE INDEX idx_order_date ON orders(date);
Check running jobs with ADMIN SHOW DDL JOBS; to find the job ID and its state.
Pause the job using ADMIN PAUSE DDL JOBS ;, which returns a success result.
Verify the pause by re-running ADMIN SHOW DDL JOBS; — the state should show as "paused."

Pausing does not cancel any currently running transactions: existing transactions continue, DDL reorganization stops, and resources are freed up for production workloads. This makes pausing safer than canceling during peak load.

Resuming a DDL Job - Commands and Timing

Once load decreases or during a maintenance window, resume the job with ADMIN RESUME DDL JOBS ;. Checking ADMIN SHOW DDL JOBS; afterward should show the state as "running." The job resumes from its last completed checkpoint rather than starting over. Many SaaS teams pause migrations during business hours and resume them during nightly maintenance windows, scheduled low-traffic periods, or region-specific off-peak hours to keep application performance consistent. If a resumed job runs into trouble, recommended steps include checking cluster health, verifying TiKV resource utilization, inspecting the DDL job queue, and canceling/recreating the job if necessary.

Real-World Use Cases

Peak traffic pause: A multi-tenant SaaS platform sees high API latency after adding an index to a large table; engineers pause the DDL job and resume it during off-peak hours instead of canceling it.
Emergency rollback without table locks: During index creation, an unexpected workload spike hits the cluster; pausing halts further resource consumption while the team investigates.
Multi-region coordination: For globally distributed deployments, DDL jobs can be paused until replication or regional traffic stabilizes, helping coordinate schema changes across regions.

Troubleshooting Common Issues

If a job ID is incorrect or the job has already completed, TiDB returns a "DDL Job Not Found" error — check ADMIN SHOW DDL JOBS; to verify. If a resumed job doesn't execute immediately, possible causes include another DDL job ahead in the queue, resource throttling, or TiKV backpressure. The same ADMIN SHOW DDL JOBS; command shows job order and state.

Best Practices

Set alerts to monitor long-running DDL job duration, always test pause/resume behavior in staging before applying it in production, and document DDL schedules so development and operations teams stay coordinated.

Conclusion

TiDB's pause/resume DDL feature gives teams control over long-running schema operations, letting them pause expensive jobs during peak workloads and resume them later without losing progress. This is especially valuable for SaaS platforms, high-traffic production systems, multi-region deployments, and large-scale data platforms.

MySQL Performance Issues: 7 Signs Your Database Needs Professional Tuning

Mafiree — Wed, 24 Jun 2026 03:56:33 +0000

MySQL performance issues rarely surface all at once. They accumulate quietly — a query that once responded in milliseconds starts taking seconds, CPU usage climbs unnoticed, and replica lag slowly worsens. By the time users or monitoring systems raise the alarm, the underlying problem has often been growing for weeks. Recognizing the early warning signs is critical to avoiding a full-blown outage.

1. Your Slow Query Log Is Being Ignored

If the same query patterns keep appearing in your slow query log week after week, the real problem isn't the queries themselves — it's the absence of a systematic process to read, triage, and fix them. Tools like pt-query-digest from Percona Toolkit can aggregate slow query entries into ranked reports, showing which query type consumes the most total execution time across the workload. The root cause is often missing composite indexes, implicit type conversions in WHERE clauses, or functions applied to indexed columns that prevent the optimizer from using them effectively — each of which requires a different resolution.

2. InnoDB Buffer Pool Hit Rate Falls Below 99%

The InnoDB buffer pool is the single most impactful memory structure in MySQL. When it is large enough, the hot working set lives in RAM; when it isn't, every cache miss becomes a disk read — and that's the fastest path to throughput collapse. A buffer pool hit rate below 99% means the server is regularly fetching data from disk. The fix isn't always adding RAM — it may involve identifying which tables or indexes are evicting hot pages, tuning innodb_buffer_pool_size, or enabling multiple buffer pool instances to reduce latch contention.

3. Replication Lag Keeps Growing Without a Clear Cause

Replication lag has multiple root causes, and each requires a different fix. Treating all lag as "the replica is slow" leads to wasted effort. The most common causes, in order of frequency, include: a single-threaded replica applier that serializes parallel writes from the source (fixed by enabling replica_parallel_workers); long-running transactions on the source that the replica must replay serially; missing indexes on the replica causing row-based replication to perform full table scans per event; network saturation between source and replica (addressed by enabling binary log transaction compression in MySQL 8.0.20+); and replica storage that cannot keep pace with the apply rate.

4. Table-Level Locks in a High-Concurrency Workload

InnoDB uses row-level locking. If table-level locks appear in an InnoDB workload, something upstream has forced a full-table lock — a DDL statement run without ALGORITHM=INPLACE, an unclosed LOCK TABLES call in application code, or a query running without an index that escalates to an implicit table lock. Persistent lock waits are an architectural signal. Remediation may include adding missing indexes, reordering transactions, or migrating DDL operations to online tools like pt-online-schema-change or gh-ost.

5. Thread Count and Mutex Waits Rise Under Normal Load

A rising thread count that isn't proportional to actual query load is a sign of contention, not capacity. Threads pile up waiting for resources — locks, buffer pool latches, or I/O — rather than actively processing work. The Performance Schema's wait event summaries can pinpoint exact culprits, such as buffer pool mutex contention (resolved by increasing innodb_buffer_pool_instances) or storage I/O latency. If the application doesn't use a connection pool, the overhead of creating and tearing down threads per request also becomes significant at scale, which can be addressed by configuring thread_cache_size or deploying a proxy layer like ProxySQL.

6. ibdata1 or Undo Tablespace Is Growing Unbounded

In configurations still using a shared system tablespace (ibdata1), or environments with large undo tablespace growth, storage consumption climbs even when actual data volume is stable. This directly impacts performance: InnoDB's write path has to manage a bloated, fragmented tablespace. The most frequent cause is long-running transactions that hold open a read view, preventing InnoDB's purge thread from cleaning up undo records. A history list length persistently above 10,000 indicates the purge thread is falling behind. The long-term resolution involves migrating to separate, truncatable undo tablespaces and rewriting the transactions responsible for holding undo records open.

7. The Query Optimizer Keeps Changing Execution Plans

If EXPLAIN output for the same logical query varies between executions — sometimes picking one index, sometimes another, sometimes doing a full scan — optimizer statistics are stale, skewed, or the sampling isn't representative of the actual data range being queried. This causes intermittent latency spikes that are difficult to reproduce on demand. The layered fix involves refreshing statistics with ANALYZE TABLE, increasing innodb_stats_persistent_sample_pages for large tables (the default of 20 is often insufficient; 200+ gives more stable estimates), adding column histograms for non-indexed columns used in WHERE clauses, and using optimizer hints to lock in the correct index for critical queries while data distribution is investigated further.

The Takeaway

These seven issues — persistent slow queries, a low buffer pool hit rate, unexplained replication lag, lock contention, thread pile-ups, tablespace bloat, and unstable execution plans — each have specific, actionable fixes. However, in most production environments they appear together, and resolving one without understanding the others leads to repetitive troubleshooting that wastes engineering time. Professional MySQL tuning requires reading the system as a whole: workload patterns, index design, memory configuration, storage behavior, replication topology, and application connection handling together.