I have a curse. My curse is curiosity. 

Here in Percona I found one person that is very “dangerous” for me, and he and I also share the same first name, well almost, Marcos Albe. 

Marcos is a smart guy with a lot of ideas, and he is not shy to share them. One day we were talking about sysbench, and he mentioned to me: you know it would be nice to have a test for joins in sysbench. I wonder why we don’t have it; it will be so useful to identify regressions in that area.
You see where this is going right? He put that in my mind, luckily I was busy with other stuff and I forgot. Until one day. That day I had to take a more in-depth look at some claims about joins performance regressions. In that moment Marcos voice comes back as an echo resonating in my mind … it would be useful to identify regressions… 

That’s it. Ok let’s do it, let’s build this joins test and see how it goes, and shut Marcos up.

I used sysbench as an engine but then I diverged a lot while implementing the test in lua. My idea was to build a simple test but got trapped in the mechanism and (never ask AI) at the end I came up with 41 different tests and a quite complicated table(s) definition. 

Let’s see both together.

Table(s) structure

In the new test we have two different sets of tables, the main table and the level table. The main table name can be renamed as you like using the usual sysbench parameter table_name, the level table name is fix levelX and for now you have queries with only one level or all of them. Already planning to make this customizable but for now it is like that.  

The main table structure is as follows:

CREATE TABLE `%s%d` (
      `id` %s,
      l1_id INT,           -- Foreign key to level1.id
      l2_id INT,           -- Foreign key to level2.id
      l3_id INT,           -- Foreign key to level3.id
      l4_id INT,           -- Foreign key to level4.id
      l5_id INT,           -- Foreign key to level5.id
      -- Numeric data types
      small_number SMALLINT,
      integer_number INT,
      myvalue BIGINT,
      decimal_number DECIMAL(10, 2),
      float_number FLOAT,
      -- String data types
      char_field CHAR(10),
      varchar_field VARCHAR(255),
      color VARCHAR(50),
      continent VARCHAR(255),
      uuid VARCHAR(36) CHARACTER SET latin1,
      uuid_bin BINARY(16),
      text_field TEXT,
      -- Date and time data types
      datetime_field DATETIME,
      timestamp_field TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
      year_field YEAR,
      -- Binary data types
      binary_field BINARY(50),
      varbinary_field VARBINARY(255),
      -- Special data types
      enum_field ENUM('active', 'inactive', 'pending'),
      set_field SET('read', 'write', 'execute', 'delete'),
      -- Boolean type
      is_active BOOLEAN DEFAULT TRUE,
      -- Spatial data type (if using GIS)
      -- point_field POINT,
      -- Constraints
      UNIQUE KEY unique_varchar (uuid),
      INDEX idx_l1_id (l1_id),
      INDEX idx_l2_id (l2_id),
      INDEX idx_l3_id (l3_id),
      INDEX idx_l4_id (l4_id),
      INDEX idx_l5_id (l5_id),
      INDEX idx_date (datetime_field),
      INDEX idx_time (timestamp_field),
      INDEX idx_enum_field (enum_field),
      INDEX idx_set_field (set_field),
      INDEX idx_year_field (year_field),
      INDEX comp_attributes(continent,enum_field,set_field),
      INDEX comp_color(color,continent,enum_field,year_field)
)

As you can see the main table has many more attributes and datatypes than the simple test existing in the basic sysbench and more than my other windmills test. 

It also has several attributes with variable length, this is to be able to emulate cases where unordered pages are present, which is the common case in production. (see also https://www.tusacentral.net/joomla/index.php/mysql-blogs/186-innodb-page-merging-and-page-splitting). To achieve that you just need to generate write traffic on the main tables, this is not available yet in the test but is not difficult to implement. As such at the moment we should consider the test executed always on Ordered pages.

Level table structure:

CREATE TABLE `%s%d` (
      `id` %s,
      continent VARCHAR(45) NOT NULL,
      parent_id BIGINT,  -- For hierarchical structure if needed
      time_accessed TIMESTAMP DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP,
      l1_id INT,
      l2_id INT,
      l3_id INT,
      l4_id INT,
      l5_id INT,    
      record_name CHAR(36),
      record_code CHAR(5),
      record_value BIGINT,
      record_status ENUM('active', 'inactive', 'pending'),
      record_priority INT NOT NULL,
      INDEX idx_country (continent),
      INDEX idx_parent_id (parent_id),
      INDEX idx_l1_id (l1_id),
      INDEX idx_l2_id (l2_id),
      INDEX idx_l3_id (l3_id),
      INDEX idx_l4_id (l4_id),
      INDEX idx_l5_id (l5_id),
      INDEX idx_time_accessed (time_accessed),
      INDEX idx_record_status (record_status),
      INDEX idx_record_priority (record_priority),
      INDEX comp_record_continent_status_priority(continent,record_status,record_priority)

In this case the table structure is much simpler (than main) and it has only one attribute (Country) which is Variable, this to reduce the possibility to have unordered pages to the minimum.

The Tests

Now combining the tables I have built the following scenarios:

·         insert_update_delete_queries

·         inner_queries

·         left_queries

·         right_queries

Pro-Tip from Celko: Avoid RIGHT JOIN whenever possible. It is mathematically equivalent to a LEFT JOIN with the tables swapped. Using only LEFT JOIN makes your queries more readable and follows the "natural" left-to-right reading order of SQL.

·         semi_join & anti join condition_queries

·         subquery_queries

Using the above I have built the following scenarios:

•    anti_join_left_join_pk 
•    anti_join_not_exists_pk 
•    conditional_join_pk 
•    inner_subquery_multi_pk 
•    left_subquery_multi_pk 
•    multi_left_forcing_order 
•    multi_left_index 
•    multi_left_pk 
•    multi_left_straight 
•    multilevel_inner_forcing_order_index 
•    multilevel_inner_index 
•    multilevel_inner_pk 
•    multilevel_inner_straight_index 
•    multi_right_forcing_order 
•    multi_right_index 
•    multi_right_pk 
•    multi_right_straight 
•    right_subquery_multi_pk 
•    semi_join_exists_pk 
•    simple_inner_forcing_order 
•    simple_inner_index 
•    simple_inner_index_GB 
•    simple_inner_pk 
•    simple_inner_pk_GB 
•    simple_inner_straight 
•    simple_left_exclude 
•    simple_left_forcing_order 
•    simple_left_index 
•    simple_left_index_GB 
•    simple_left_pk 
•    simple_left_pk_GB 
•    simple_left_straight 
•    simple_right_forcing_order 
•    simple_right_index 
•    simple_right_index_GB 
•    simple_right_pk 
•    simple_right_pk_GB 
•    simple_right_straight 
•    update_multi_inner_join_pk 
•    update_multi_left_join_pk 
•    update_multi_right_join_pk

Where simple means only one level of joins while multi is all the five levels. 

If you want to review them you can easily do it reviewing the file in github (https://github.com/Tusamarco/sysbench/blob/master/src/lua/joins/join_queries.lua)

Please keep in mind a couple of things:

1. This is just a start; I do not pretend to cover all and I am open to add whatever it will be identified as useful.

2. You are encouraged to test and modify any part of the code, query, table structure and share your changes, the more we work on this together the better it will be. 

How to use it

First clone the repo as usual, build sysbench and install it, instructions here(https://github.com/Tusamarco/sysbench/tree/master?tab=readme-ov-file#build-and-install). 

Unless you do it manually the tests in the src/lua/… will not be copy over if you want you can modify the file https://github.com/Tusamarco/sysbench/blob/master/src/lua/Makefile.am to have them installed with make install.

As usual the first step is to create and populate the schema:

sysbench /opt/sysbench/src/lua/joins/oltp_read_write.lua  --mysql-host=<ip> --mysql-port=<port> --mysql-user=<user> --mysql-password=<pw> --mysql-db=joins --db-driver=mysql  --report-interval=1  --table_name=main  --tables=15 --table_size=100000 --threads=15  prepare

This will create 15 main tables and 5 levelX tables, both with 100000 rows.

I suggest you create a schema with a decent number of rows, in general the dimension of all rows in a table should exceed the allocated memory buffer. So, if you have a 10GB Innodb buffer pool, it would be nice to have each table to be at least 15GB.

Always do a warmup, that will fill the BP with data, most likely from the last main table, but this is fine, you just want the BP to be hot. 

Once done you can run the tests. 

To evaluate the join execution, I think that the most immediate metric is the execution time. Given that when you run the test you should set –time=0 –events=X where X is the number of events you want to execute. 

How many queries will an event have? Easy, if you define it as a parameter, by default all the joins are set to 0, which means disable and you need to enable it. 

For instance if you want to execute simple_inner_pk test, you need to set --simple_inner_pk=1 where 1 is the number of occurrence, so if you set it to 1 you have one event one query, if you set --simple_inner_pk=2 you have that each event generate 2 queries… and so on. 

Given the time taken is captured for the whole test, my advice is to keep the occurrence to 1 and eventually check/test how the platform scales by concurrent threads.

To do so, you need to set the number of events and threads to the same value. 

See a full example below:

sysbench /opt/sysbench/src/lua/joins/oltp_read_write.lua  --mysql-host=<ip> --mysql-port=<port> --mysql-user=<user> --mysql-password=<pw> --mysql-db=joins --db-driver=mysql  --skip_trx=off --report-interval=1  --histogram --table_name=main  --stats_format=csv --tables=15 --table_size=1000000 --time=0 --multilevel_inner_pk=1 --events=5 –threads=5 run

This will record the time taken to execute 5 queries in parallel.

While you can enable more than one test at the same time, to me this is not useful given it will be not clear who is taking longer unless you do not keep track of that in other ways, like slow query log or performance schema.  

How to modify it

You want to modify the queries or table structure?

Do it! Just modify the file joins_queries.lua. There you will find the table definition for the main and level table.
The only thing to be careful of is to change the variables in the load_global_variables() to match your changes and of course you will need to modify the initialize_values_X for main and level table.

If instead you only want to change the queries you have two ways.

1. Add your query to an already existing map 

2. Create a new map

Method 1. 

Say you want to add a query for the left joins. Identify the left_queries = {…} map and add the query you want to add:

["my_custom_left_query"] = [[SELECT m.continent,year_field, m.enum_field, level1.record_value l1
FROM %s%u as m
LEFT JOIN level1 ON m.l1_id = level1.id and m.enum_field = '%s'
WHERE m.continent = '%s'
ORDER BY m.year_field DESC, l1 DESC
LIMIT 100;]],

Here I am not adding anything special, so I just need to add a pair of [KEY] = <SELECT> using the same attributes. 

In case you need to change the attributes, then you need to create a case in the function that fill the query. The function is: function execute_joins(join_name) in joins/oltp_common.lua. 

For instance if I want to have my custom query filtering by the attribute color in the main table, I can do:

["my_custom_left_query"] = [[SELECT m.continent,year_field, m.enum_field, level1.record_value l1
FROM %s%u as m
LEFT JOIN level1 ON m.l1_id = level1.id and m.enum_field = '%s'
WHERE m.color = '%s'
ORDER BY m.year_field DESC, l1 DESC
LIMIT 100;]],

In the function I can add a case like:

  if join_name:find("my_custom_left") then
      for i = 1, sysbench.opt[join_name] do
         local tnum = get_table_num()
         query = string.format(query_map[join_name .. "_query"], sysbench.opt.table_name, tnum, get_record_status(), get_color())
         -- print("DEBUG JOIN QUERY B: " .. query .." Join Name: " .. join_name)
         con:query(query)
      end

Done!

Method 2

Just respect the rules as above but instead of using the existing map, you can create your own and then add it to the list query_map = mergeMultiple. As simple as that.

How to help

As said before, this is far from a perfect or exhaustive test, it is a start and the more feedback I get the more I will be able to improve it. 

I would really like to get PR or FR with specific tasks, the closer they are to real life cases the better.

In the meantime, I wish you all happy testing.

References

Joins sysbench code

MySQL 8.4: JOIN Clause – The definitive reference for INNER, LEFT, RIGHT, and CROSS joins.

Outer Join Optimization – Explains how the optimizer handles LEFT JOIN and RIGHT JOIN and how to write efficient predicates.

Nested-Loop Join Algorithm – For those interested in the "under the hood" execution of joins in MySQL.

Index Merge Optimization – Crucial for understanding how indexes are used (or not used) during complex joins.

"What happens when you run a 10 TB MySQL database on Kubernetes?" 

That's the question many of our customers and users asked and honestly, we were extremely curious ourselves.

So, we ditched the weekend plans, rolled up our sleeves, and jumped down the rabbit hole. What we found was far more challenging (and perhaps a bit more "psychedelic") than expected. We spent days rigorously testing the Percona Operator for PXC at massive scale. 

This blog post distills all our findings into the most critical, actionable advice you need to ensure your high-scale environment not only survives but operates reliably.

If you are lazy and prefer to watch a video, here is the presentation I did at the MySQL Belgian days in January 2026.

First of all, what environment did we use? 

Well forget about 4Gb ram or using the millesimal set on CPUs. You need heavy stuff, you need power, you need the mega-super-duper tank not the small bike, well if you can pay for it of course. Anyhow we have tested on something that is not small and not too huge as well, all in AWS/EKS.

The AWS instance type chosen for the testing environment was m7i.16xlarge, combined with an io2 storage class provisioned at 80,000 IOPS.

This configuration was selected because:

• The m7i.16xlarge is a general-purpose instance with custom 4th Generation Intel Xeon Scalable processors, offering 64 vCPUs and 256 GiB of high-bandwidth DDR5 memory. This robust computational and memory capacity ensures that CPU or memory are unlikely to be limiting factors during initial testing of a large-scale, high-performance MySQL cluster on EKS.
• Its 20 Gbps dedicated EBS bandwidth and up to 80,000 maximum IOPS to EBS perfectly complement the high-performance io2 Block Express storage volume.
• By provisioning the storage to the maximum rated IOPS of the underlying instance type, the aim was to create a stress-testing environment that isolates potential performance constraints in the MySQL operator, Percona architecture, Kubernetes storage mechanisms (CSI/EBS), or database configuration, rather than hitting a hardware-imposed ceiling. This allows for a precise analysis of operational overheads and bottlenecks when the I/O subsystem is not the weakest link, which is critical when scaling a database to 10 TB of data.

Did it go as we would like to? Well long story short, not in full but I was expecting worse … into the rabbit hole we go.

We fully validated the cluster's ability to scale, but here's what surprised (and alarmed) us the most:

• The 16-Hour Naptime MTTR (Mean Time To Recover): Full cluster recovery time for the 10 TB dataset clocked in at an eye-watering 960 minutes (16 hours)! If your business requires a guaranteed 1-hour Recovery Time Objective (RTO), you're going to have a very awkward conversation with your boss.

The rebuild operation from cluster crash was executed recovering the first node from snapshot then, letting the cluster recover using internal recovery mechanism.

• The Unruly Cloud Storage: We provisioned high-performance AWS EBS volumes (80,000 IOPS), yet we hit a wall due to Galera Flow Control. The root cause? One node's volume had a 21% performance degradation compared to the others. The entire synchronous cluster was choked by the single slowest volume.

• The Starved Recovery Container: We discovered the restore process's most time-consuming phase the single-threaded redolog apply was bottlenecked because the default recovery container only gets a laughably default memory allocation (around 100 MB). This process consumed 92% of the total restore time.

The Safe Operation Playbook: Critical Fixes

You can tame the data monster, but you have to stop using the default settings. Your focus must shift to I/O consistency, recovery optimization, and mandatory zero-downtime practices.

1. Stop Starving Your Recovery Container

The single most effective action to improve your RTO is giving the restore container more memory to efficiently process the redolog.

• The Fix: You must override the inadequate defaults in your recovery.yaml specification.
• Recommendation: Set memory: 4Gi and cpu: 2000m (2 full cores). This directly reduces the time spent on the single-threaded redolog apply phase, chipping away at that painful 16-hour recovery window. This will alleviate and not resolve the situation though. 
• Redefine Recovery SLO: Since the Operator waits for full cluster recovery before HAProxy serves traffic, you must redefine your internal service goal to Minimal High Availability (HA) (two out of three nodes synced), you can achieve it increasing the size of the cluster by steps, so first one node then add the second and so on. This allows you to claim service is resumed in hours (or less, if optimized) rather than 16 hours.

2. The Cloud Storage Problem: EBS is Not Trustable

For critical, high-scale PXC clusters, the testing proved that high-cost abstracted cloud volumes (EBS) are too unpredictable due to the risk of I/O variability.

• The Problem: I/O variance is a single point of failure that triggers Galera Flow Control. Because PXC is synchronous, if one node is 21% slower, the entire cluster is 21% slower.
• The Long-Term Solution: For mission-critical 10 TB systems, mandate a review for dedicated storage solutions. Look for cloud alternatives that offer strict I/O guarantees, such as dedicated NVMe instances, to bypass the performance abstraction layer and ensure every node performs equally. Or opt for direct attach (fiberchannel) solutions with a SAN tune to prevent volume overlaps. 

3.  Zero-Downtime is Mandatory (Say No to Direct SQL DDL)

On a Percona XtraDB Cluster, DDL is replicated using Total Order Isolation (TOI). This means that direct ALTER TABLE operations cause a full, unacceptable table lock, resulting in downtime.

• The Rule: Use PTOSC (Percona Online Schema Change) for all DDL, including simple index operations. It takes significantly longer, but it is the only way to avoid service interruption.
• The Capacity Warning: PTOSC temporarily duplicates the entire target table. Ensure you have meticulous capacity planning for temporary disk space to prevent a cluster-wide storage saturation crash!
• Large DML: Break all large UPDATE and DELETE operations into small, iterative batches (chunks) to minimize the scope and duration of transactional locks.

Hot Data Subset vs. The Whole Barn

The most fascinating performance finding was how much the application's access pattern affects performance at scale. This dictates how well your InnoDB Buffer Pool can save you from a slow disk.

Remember that the tests used a 50/50 read/write workload. Even when operating on a "hot" subset in the buffer pool, the persistent write and purge operations still required disk I/O, which likely reduced the visible performance difference (delta) between the hot subset and uniform access patterns. This delta would be larger with a higher percentage of read operations. 

The Bottom Line is that your 10 TB cluster is faster if your application keeps queries focused on the "hot" data subset. Uniform access patterns force the system to pay the full 10 TB I/O cost, every time. 

Actionable Summary for the Team

1. Recovery Override: Set memory: 4Gi in the recovery spec to reduce the negative impact of the 92% redolog apply bottleneck.
2. DDL Enforcement: PTOSC is mandatory for all schema changes. Batch large DML.
3. Storage Fix: Do not trust EBS/abstracted storage for 10 TB. Plan an architectural shift toward guaranteed I/O (e.g., dedicated NVMe instances).
4. Data Mobility: Standardize on MyDumper/Loader for all bulk operations.
5. Data distribution/access: If You Have a Hot Data Subset, Offload the Rest. The better performance seen with the "hot" data subset (Pareto) is your green light!
   Any data your application accesses uniformly (or rarely) is only increasing your recovery time and operational cost.
   You should treat the remaining data as a candidate for immediate migration to an OLAP container or data lake, ensuring your PXC cluster remains lean and fast.

Conclusions

Ultimately, our deep dive proved that running Percona XtraDB Cluster on Kubernetes with a 10 TB dataset is less about raw scalability and more about operational rigor and resource tuning. 

The cluster is fundamentally resilient, but its success hinges on bypassing the hidden pitfalls like:

• the unpredictable nature of abstract cloud I/O
• the crippling cost of default recovery settings
• the unacceptable downtime from standard DDL practices.

However we also identify some areas for improvements like in the case of full cluster recovery, we will work on that. 

The key takeaway is simple: default configurations are for small data, large datasets require deliberate engineering.

Notes:

*  In this context means: Mean Time To Recover / Restore (The "User" Metric). Which is the average time from the start of the outage until the system is fully back online and usable for the customer. Includes: Detection time + Response time + Repair time + Testing time.