Best Practices for Migrating OpenTSDB Applications to TDengine

As a distributed, scalable, distributed time-series database platform based on HBase, and thanks to its first-mover advantage, OpenTSDB is widely used for monitoring in DevOps. However, as new technologies like cloud computing, microservices, and containerization technology has developed rapidly, Enterprise-level services are becoming more and more diverse and the architecture is becoming more complex.

As a result, as a DevOps backend for monitoring, OpenTSDB is plagued by performance issues and delayed feature upgrades. This has resulted in increased application deployment costs and reduced operational efficiency. These problems become increasingly severe as the system tries to scale up.

To meet the fast-growing IoT big data market and technical needs, TAOSData developed an innovative big-data processing product, TDengine.

After learning the advantages of many traditional relational databases and NoSQL databases, stream computing engines, and message queues, TDengine has its unique benefits in time-series big data processing. TDengine can effectively solve the problems currently encountered by OpenTSDB.

Compared with OpenTSDB, TDengine has the following distinctive features.

• Data writing and querying performance far exceeds that of OpenTSDB.
• Efficient compression mechanism for time-series data, which compresses to less than 1/5 of the storage space, on disk.
• The installation and deployment are straightforward. A single installation package can complete the installation and deployment and does not rely on other third-party software. The entire installation and deployment process takes a few seconds.
• The built-in functions cover all of OpenTSDB's query functions and TDengine supports more time-series data query functions, scalar functions, and aggregation functions. TDengine also supports advanced query functions such as multiple time-window aggregations, join query, expression operation, multiple group aggregation, user-defined sorting, and user-defined functions. With a SQL-like query language, querying is more straightforward and has no learning cost.
• Supports up to 128 tags, with a total tag length of 16 KB.
• In addition to the REST interface, it also provides interfaces to Java, Python, C, Rust, Go, C# and other languages. Its supports a variety of enterprise-class standard connector protocols such as JDBC.

Migrating applications originally running on OpenTSDB to TDengine, effectively reduces compute and storage resource consumption and the number of deployed servers. It also significantly reduces operation and maintenance costs, makes operation and maintenance management more straightforward and more accessible, and considerably reduces the total cost of ownership. Like OpenTSDB, TDengine has also been open-sourced. Both the stand-alone version and the cluster version are open-sourced and there is no need to be concerned about the vendor-lock problem.

We will explain how to migrate OpenTSDB applications to TDengine quickly, securely, and reliably without coding, using the most typical DevOps scenarios. Subsequent chapters will go into more depth to facilitate migration for non-DevOps systems.

DevOps Application Quick Migration

1. Typical Application Scenarios

The following figure (Figure 1) shows the system's overall architecture for a typical DevOps application scenario.

Figure 1. Typical architecture in a DevOps scenario

In this application scenario, there are Agent tools deployed in the application environment to collect machine metrics, network metrics, and application metrics. There are also data collectors to aggregate information collected by agents, systems for persistent data storage and management, and tools for data visualization (e.g., Grafana, etc.).

The agents deployed in the application nodes are responsible for providing operational metrics from different sources to collectd/Statsd. And collectd/StatsD is accountable for pushing the aggregated data to the OpenTSDB cluster system and then visualizing the data using the visualization kanban board software, Grafana.

2. Migration Services

• TDengine installation and deployment

First of all, please install TDengine. Download the latest stable version of TDengine from the official website and install it.

Note that once the installation is complete, do not start the taosd service before properly configuring the parameters.

• Adjusting the data collector configuration

TDengine version 2.4 and later version includes taosAdapter. taosAdapter is a stateless, rapidly elastic, and scalable component. taosAdapter supports Influxdb's Line Protocol and OpenTSDB's telnet/JSON writing protocol specification, providing rich data access capabilities, effectively saving user migration costs and reducing the difficulty of user migration.

Users can flexibly deploy taosAdapter instances, based on their requirements, to improve data writing throughput and provide guarantees for data writes in different application scenarios.

Through taosAdapter, users can directly write the data collected by collectd or StatsD to TDengine to achieve easy, convenient and seamless migration in application scenarios. taosAdapter also supports Telegraf, Icinga, TCollector, and node_exporter data. For more details, please refer to taosAdapter.

If using collectd, modify the configuration file in its default location /etc/collectd/collectd.conf to point to the IP address and port of the node where to deploy taosAdapter. For example, assuming the taosAdapter IP address is 192.168.1.130 and port 6046, configure it as follows.

LoadPlugin write_tsdb
<Plugin write_tsdb>
  <Node>
    Host "192.168.1.130" Port "6046" HostTags "status=production" StoreRates
    false AlwaysAppendDS false
  </Node>
</Plugin>

You can use collectd and push the data to taosAdapter utilizing the write_tsdb plugin. taosAdapter will call the API to write the data to TDengine. If you are using StatsD, adjust the profile information accordingly.

• Tuning the Dashboard system

After writing the data to TDengine, you can configure Grafana to visualize the data written to TDengine. To obtain and use the Grafana plugin provided by TDengine, please refer to Links to other tools.

TDengine provides two sets of Dashboard templates by default, and users only need to import the templates from the Grafana directory into Grafana to activate their use.

Importing Grafana Templates Figure 2.

With the above steps completed, you have finished replacing OpenTSDB with TDengine. You can see that the whole process is straightforward, there is no need to write any code, and only some configuration files need to be changed.

3. Post-migration architecture

After completing the migration, the figure below (Figure 3) shows the system's overall architecture. The whole process of the acquisition side, the data writing, and the monitoring and presentation side are all kept stable. There are a few configuration adjustments, which do not involve any critical changes or alterations. Migrating to TDengine from OpenTSDB leads to powerful processing power and query performance.

In most DevOps scenarios, if you have a small OpenTSDB cluster (3 or fewer nodes) which provides storage and data persistence layer in addition to query capability, you can safely replace OpenTSDB with TDengine. TDengine will save compute and storage resources. With the same compute resource allocation, a single TDengine can meet the service capacity provided by 3 to 5 OpenTSDB nodes. TDengine clustering may be required depending on the scale of the application.

Figure 3. System architecture after migration

The following chapters provide a more comprehensive and in-depth look at the advanced topics of migrating an OpenTSDB application to TDengine. This will be useful if your application is particularly complex and is not a DevOps application.

Migration evaluation and strategy for other scenarios

1. Differences between TDengine and OpenTSDB

This chapter describes the differences between OpenTSDB and TDengine at the system functionality level. After reading this chapter, you can fully evaluate whether you can migrate some complex OpenTSDB-based applications to TDengine, and what you should pay attention to after migration.

TDengine currently only supports Grafana for visual kanban rendering, so if your application uses front-end kanban boards other than Grafana (e.g., TSDash, Status Wolf, etc.) you cannot directly migrate those front-end kanbans to TDengine. The front-end kanban will need to be ported to Grafana to work correctly.

TDengine version 2.3.0.x only supports collectd and StatsD as data collection and aggregation software but future versions will provide support for more data collection and aggregation software in the future. If you use other data aggregators on the collection side, your application needs to be ported to these two data aggregation systems to write data correctly. In addition to the two data aggregator software protocols mentioned above, TDengine also supports writing data directly via InfluxDB's line protocol and OpenTSDB's data writing protocol, JSON format. You can rewrite the logic on the data push side to write data using the line protocols supported by TDengine.

In addition, if your application uses the following features of OpenTSDB, you need to take into account the following considerations before migrating your application to TDengine.

1. /api/stats: If your application uses this feature to monitor the service status of OpenTSDB, and you have built the relevant logic to link the processing in your application, then this part of the status reading and fetching logic needs to be re-adapted to TDengine. TDengine provides a new mechanism for handling cluster state monitoring to meet the monitoring and maintenance needs of your application.
2. /api/tree: If you rely on this feature of OpenTSDB for the hierarchical organization and maintenance of timelines, you cannot migrate it directly to TDengine, which uses a database -> super table -> sub-table hierarchy to organize and maintain timelines, with all timelines belonging to the same super table in the same system hierarchy. But it is possible to simulate a logical multi-level structure of the application through the unique construction of different tag values.
3. Rollup And PreAggregates: The use of Rollup and PreAggregates requires the application to decide where to access the Rollup results and, in some scenarios, to access the actual results. The opacity of this structure makes the application processing logic extraordinarily complex and not portable at all. While TDengine does not currently support automatic downsampling of multiple timelines and preaggregation (for a range of periods), thanks to its high-performance query processing logic, it can provide very high-performance query responses without relying on Rollup and preaggregation (for a range of periods). This makes your application query processing logic straightforward and simple.
4. Rate: TDengine provides two functions to calculate the rate of change of values, namely Derivative (the result is consistent with the Derivative behavior of InfluxDB) and IRate (the result is compatible with the IRate function in Prometheus). However, the results of these two functions are slightly different from that of Rate. But the TDengine functions are more powerful. In addition, TDengine supports all the calculation functions provided by OpenTSDB. TDengine's query functions are much more powerful than those supported by OpenTSDB, which can significantly simplify the processing logic of your application.

With the above introduction, we believe you should be able to understand the changes brought about by the migration of OpenTSDB to TDengine. And this information will also help you correctly determine whether you should migrate your application to TDengine to experience the powerful and convenient time-series data processing capability provided by TDengine.

2. Migration strategy suggestion

OpenTSDB-based system migration involves data schema design, system scale estimation, data write transformation, data streaming, and application changes. The two systems should run in parallel for a while and then the historical data should be migrated to TDengine if your application has some functions that strongly depend on the above OpenTSDB features and you do not want to stop using them. You can also consider keeping the original OpenTSDB system running while using TDengine to provide the primary services.

Data model design

On the one hand, TDengine requires a strict schema definition for its incoming data. On the other hand, the data model of TDengine is richer than that of OpenTSDB, and the multi-valued model is compatible with all single-valued model building requirements.

Let us now assume a DevOps scenario where we use collectd to collect the underlying metrics of the device, including memory, swap, disk, etc. The schema in OpenTSDB is as follows.

| metric | value name | type | tag1 | tag2 | tag3 | tag4 | tag5 | | ---- | -------------- | ------ | ------ | ---- | ----------- | -------------------- | --------- | ------ | | 1 | memory | value | double | host | memory_type | memory_type_instance | source | n/a | | 2 | swap | value | double | host | swap_type | swap_type_instance | source | n/a | | 3 | disk | value | double | host | disk_point | disk_instance | disk_type | source |

TDengine requires the data stored to have a data schema, i.e., you need to create a super table and specify the schema of the super table before writing the data. For data schema creation, you have two ways to do this:

1. Take advantage of TDengine's native data writing support for OpenTSDB by calling the TDengine API to write (text line or JSON format) and automate the creation of single-value models. This approach does not require significant adjustments to the data writing application, nor does it require converting the written data format.

At the C level, TDengine provides the taos_schemaless_insert() function to write data in OpenTSDB format directly (in early version this function was named taos_insert_lines()). Please refer to the sample code schemaless.c in the installation package directory as reference.

(2) Based on a thorough understanding of TDengine's data model, establish a mapping between OpenTSDB and TDengine's data model. Considering that OpenTSDB is a single-value mapping model, we recommended using the single-value model in TDengine for simplicity. But keep in mind that TDengine supports both multi-value and single-value models.

• Single-valued model.

The steps are as follows:

• Use the name of the metrics as the name of the TDengine super table
• Build with two basic data columns - timestamp and value. The label of the super table is equivalent to the label information of the metrics, and the number of labels is equal to the number of labels of the metrics.
• The names of sub-tables are named with fixed rules: metric + '_' + tags1_value + '_' + tag2_value + '_' + tag3_value ... as the sub-table name.

Create 3 super tables in TDengine.

create stable memory(ts timestamp, val float) tags(host binary(12), memory_type binary(20), memory_type_instance binary(20), source binary(20)) ;
create stable swap(ts timestamp, val double) tags(host binary(12), swap_type binary(20), swap_type_binary binary(20), source binary(20));
create stable disk(ts timestamp, val double) tags(host binary(12), disk_point binary(20), disk_instance binary(20), disk_type binary(20), source binary(20));

For sub-tables use dynamic table creation as shown below.

insert into memory_vm130_memory_buffered_collectd using memory tags('vm130', 'memory', ' buffer', 'collectd') values(1632979445, 3.0656);

The final system will have about 340 sub-tables and three super-tables. Note that if the use of concatenated tagged values causes the sub-table names to exceed the system limit (191 bytes), then some encoding (e.g., MD5) needs to be used to convert them to an acceptable length.

• Multi-value model

Ideally you should take advantage of TDengine's multi-value modeling capabilities. In that case, you first need to meet the requirement that different collection quantities have the same collection frequency and can reach the data write side simultaneously via a message queue, thus ensuring writing multiple metrics at once, using SQL statements. The metric's name is used as the name of the super table to create a multi-column model of data that has the same collection frequency and can arrive simultaneously. The sub-tables are named using a fixed rule. Each of the above metrics contains only one measurement value, so converting it into a multi-value model is impossible.

Data triage and application adaptation

Subscribe to the message queue and start writing data to TDengine.

After data has been written for a while, you can use SQL statements to check whether the amount of data written meets the expected writing requirements. Use the following SQL statement to count the amount of data.

select count(*) from memory

After completing the query, if the data written does not differ from what is expected and there are no abnormal error messages from the writing program itself, you can confirm that the written data is complete and valid.

TDengine does not support querying, or data fetching using the OpenTSDB query syntax but does provide a counterpart for each of the OpenTSDB queries. The corresponding query processing can be adapted and applied in a manner obtained by examining Appendix 1. To fully understand the types of queries supported by TDengine, refer to the TDengine user manual.

TDengine supports the standard JDBC 3.0 interface for manipulating databases, but you can also use other types of high-level language connectors for querying and reading data to suit your application. Please read the user manual for specific operations and usage.

Historical Data Migration

1. Use the tool to migrate data automatically

To facilitate historical data migration, we provide a plug-in for the data synchronization tool DataX, which can automatically write data into TDengine.The automatic data migration of DataX can only support the data migration process of a single value model.

After migrating via DataX, we found that we can significantly improve the efficiency of migrating historical data by starting multiple processes and migrating numerous metrics simultaneously. The following are some records of the migration process. We provide these as a reference for application migration.

| Number of datax instances (number of concurrent processes) | Migration record speed (pieces/second) | | ----------------------------- | ------------------- -- | | 1 | About 139,000 | | 2 | About 218,000 | | 3 | About 249,000 | | 5 | About 295,000 | | 10 | About 330,000 |

(Note: The test data comes from a single-node Intel(R) Core(TM) i7-10700 CPU@2.90GHz 16-core 64G hardware device, the channel and batchSize are 8 and 1000 respectively, and each record contains 10 tags)

2. Manual data migration

Suppose you need to use the multi-value model for data writing. In that case, you need to develop a tool to export data from OpenTSDB, confirm which timelines can be merged and imported into the same timeline, and then pass the time to import simultaneously through the SQL statement—written to the database.

Manual migration of data requires attention to the following two issues:

1. When storing the exported data on the disk, the disk needs to have enough storage space to accommodate the exported data files fully. To avoid running out of disk space, you can adopt a partial import mode in which you preferentially export the timelines belonging to the same super table and then only those files are imported into TDengine.

2. Under the full load of the system, if there are enough remaining computing and IO resources, establish a multi-threaded import to maximize the efficiency of data migration. Considering the vast load that data parsing brings to the CPU, it is necessary to control the maximum number of parallel tasks to avoid overloading the system when importing historical data.

Due to the ease of operation of TDengine itself, there is no need to perform index maintenance and data format change processing in the entire process. The whole process only needs to be executed sequentially.

While importing historical data into TDengine, the two systems should run simultaneously. Once all the data is migrated, switch the query request to TDengine to achieve seamless application switching.

Appendix 1: OpenTSDB query function correspondence table

Avg

Equivalent function: avg

Example:

SELECT avg(val) FROM (SELECT first(val) FROM super_table WHERE ts >= startTime and ts <= endTime INTERVAL(20s) Fill(linear)) INTERVAL(20s)

Remarks:

1. The value in Interval needs to be the same as the interval value in the outer query.
2. Interpolation processing in TDengine uses subqueries to assist in completion. As shown above, it is enough to specify the interpolation type in the inner query. Since OpenTSDB uses linear interpolation, use fill(linear) to declare the interpolation type in TDengine. Some of the functions mentioned below have exactly the same interpolation calculation requirements.
3. The parameter 20s in Interval indicates that the inner query will generate results according to a time window of 20 seconds. In an actual query, it needs to adjust to the time interval between different records. It ensures that interpolation results are equivalent to the original data.
4. Due to the particular interpolation strategy and mechanism of OpenTSDB i.e. interpolation followed by aggregate calculation, it is impossible for the results to be completely consistent with those of TDengine. But in the case of downsampling (Downsample), TDengine and OpenTSDB can obtain consistent results (since OpenTSDB performs aggregation and downsampling queries).

Count

Equivalent function: count

Example:

select count(\*) from super_table_name;

Dev

Equivalent function: stddev

Example:

Select stddev(val) from table_name

Estimated percentiles

Equivalent function: apercentile

Example:

Select apercentile(col1, 50, “t-digest”) from table_name

Remark:

1. When calculating estimate percentiles, OpenTSDB uses the t-digest algorithm by default. In order to obtain the same calculation results in TDengine, the algorithm used needs to be specified in the apercentile() function. TDengine can support two different percentile calculation algorithms named "default" and "t-digest" respectively.

First

Equivalent function: first

Example:

Select first(col1) from table_name

Last

Equivalent function: last

Example:

Select last(col1) from table_name

Max

Equivalent function: max

Example:

Select max(value) from (select first(val) value from table_name interval(10s) fill(linear)) interval(10s)

Note: The Max function requires interpolation for the reasons described above.

Min

Equivalent function: min

Example:

Select min(value) from (select first(val) value from table_name interval(10s) fill(linear)) interval(10s);

MinMax

Equivalent function: max

Select max(val) from table_name

Note: This function has no interpolation requirements, so it can be directly calculated.

MimMin

Equivalent function: min

Select min(val) from table_name

Note: This function has no interpolation requirements, so it can be directly calculated.

Percentile

Equivalent function: percentile

Remark:

Sum

Equivalent function: sum

Select max(value) from (select first(val) value from table_name interval(10s) fill(linear)) interval(10s)

Note: This function has no interpolation requirements, so it can be directly calculated.

Zimsum

Equivalent function: sum

Select sum(val) from table_name

Note: This function has no interpolation requirements, so it can be directly calculated.

Complete example:

// OpenTSDB query JSON
query = {
"start": 1510560000,
"end": 1515000009,
"queries": [{
"aggregator": "count",
"metric": "cpu.usage_user",
}]
}

// Equivalent query SQL:
SELECT count(*)
FROM `cpu.usage_user`
WHERE ts>=1510560000 AND ts<=1515000009

Appendix 2: Resource Estimation Methodology

Data generation environment

We still use the hypothetical environment from Chapter 4. There are three measurements. Respectively: the data writing rate of temperature and humidity is one record every 5 seconds, and the timeline is 100,000. The writing rate of air pollution is one record every 10 seconds, the timeline is 10,000, and the query request frequency is 500 QPS.

Storage resource estimation

Assuming that the number of sensor devices that generate data and need to be stored is n, the frequency of data generation is t per second, and the length of each record is L bytes, the scale of data generated per day is 86400 * n * t * L bytes. Assuming the compression ratio is C, the daily data size is (86400 * n * t * L)/C bytes. The storage resources are estimated to accommodate the data scale for 1.5 years. In the production environment, the compression ratio C of TDengine is generally between 5 and 7. With additional 20% redundancy, you can calculate the required storage resources:

(86400 * n * t * L) * (365 * 1.5) * (1+20%)/C

Substituting in the above formula, the raw data generated every year is 11.8TB without considering the label information. Note that tag information is associated with each timeline in TDengine, not every record. The amount of data to be recorded is somewhat reduced relative to the generated data, and label data can be ignored as a whole. Assuming a compression ratio of 5, the size of the retained data ends up being 2.56 TB.

Storage Device Selection Considerations

A disk with better random read performance, such as an SSD, improves the system's query performance and improves the query response performance of the whole system. To obtain better query performance, the performance index of the single-threaded random read IOPS of the hard disk device should not be lower than 1000, and it is better to reach 5000 IOPS or more. We recommend using fio utility software to evaluate the running performance (please refer to Appendix 1 for specific usage) for the random IO read of the current device to confirm whether it can meet the requirements of random read of large files.

Hard disk writing performance has little effect on TDengine. The TDengine writing process adopts the append write mode, so as long as it has good sequential write performance, both SAS hard disks and SSDs in the general sense can well meet TDengine's requirements for disk write performance.

Computational resource estimates

Due to the characteristics of IoT data, when the frequency of data generation is consistent, the writing process of TDengine maintains a relatively fixed amount of resource consumption (computing and storage). According to the TDengine Operation and Maintenance Guide description, the system consumes less than 1 CPU core at 22,000 writes per second.

In estimating the CPU resources consumed by the query, assuming that the application requires the database to provide 10,000 QPS, the CPU time consumed by each query is about 1 ms. The query provided by each core per second is 1,000 QPS, which satisfies 10,000 QPS. The query request requires at least 10 cores. For the system as a whole system to have less than 50% CPU load, the entire cluster needs twice as many cores i.e. 20 cores.

Memory resource estimation

The database allocates 16MB*3 buffer memory for each Vnode by default. If the cluster system includes 22 CPU cores, TDengine will create 22 Vnodes (virtual nodes) by default. Each Vnode contains 1000 tables, which is more than enough to accommodate all the tables in our hypothetical scenario. Then it takes about 1.5 hours to write a block, which triggers persistence to disk without requiring any adjustment. A total of 22 Vnodes require about 1GB of memory cache. Considering the memory needed for the query, assuming that the memory overhead of each query is about 50MB, the memory required for 500 queries concurrently is about 25GB.

In summary, using a single 16-core 32GB machine or a cluster of 2 8-core 16GB machines is enough.

Appendix 3: Cluster Deployment and Startup

TDengine provides a wealth of help documents to explain many aspects of cluster installation and deployment. Here is the list of documents for your reference.

Cluster Deployment

The first is TDengine installation. Download the latest stable version of TDengine from the official website, and install it.

Note that once the installation is complete, do not immediately start the taosd service, but start it after correctly configuring the parameters.

Set running parameters and start the service

To ensure that the system can obtain the necessary information for regular operation. Please set the following vital parameters correctly on the server:

FQDN, firstEp, secondEP, dataDir, logDir, tmpDir, serverPort. For the specific meaning and setting requirements of each parameter, please refer to the document "TDengine Cluster Installation and Management"

Follow the same steps to set parameters on the other nodes, start the taosd service, and then add Dnodes to the cluster.

Finally, start taos and execute the show dnodes command. If you can see all the nodes that have joined the cluster, the cluster building process was successfully completed. For specific operation procedures and precautions, please refer to the document "Cluster Deployment".

Appendix 4: Super Table Names

Since OpenTSDB's metric name has a dot (".") in it, for example, a metric with a name like "cpu.usage_user", the dot has a special meaning in TDengine and is a separator used to separate database and table names. TDengine also provides "escape" characters to allow users to use keywords or special separators (e.g., dots) in (super)table names. To use special characters, enclose the table name in escape characters, e.g.: cpu.usage_user. It is a valid (super) table name.

Appendix 5: Reference Articles

1. Using TDengine + collectd/StatsD + Grafana to quickly build an IT operation and maintenance monitoring system
2. Write collected data directly to TDengine through collectd