This page contains information about Cloud Bigtable schema design. Before you read this page, you should be familiar with the overview of Cloud Bigtable. The following topics are covered on this page:
- General concepts: Basic concepts to keep in mind as you design your database.
- Best practices: Design guidelines that apply to most use cases, broken down by table component.
- Special use cases: Recommendations for some specific use cases and data patterns.
Designing a Cloud Bigtable schema is different than designing a schema for a relational database. In Cloud Bigtable, a schema is a blueprint or model of a table, including the structure of the following table components:
- Row keys
- Column families, including their garbage-collection policies
The following general concepts apply to Cloud Bigtable schema design:
- Cloud Bigtable is a key/value store, not a relational store. It does not support joins, and transactions are supported only within a single row.
- Each table has only one index, the row key. There are no secondary indices. Each row key must be unique.
- Rows are sorted lexicographically by row key, from the lowest to the highest byte string. Row keys are sorted in big-endian byte order (sometimes called network byte order), the binary equivalent of alphabetical order.
- Column families are not stored in any specific order.
- Columns are grouped by column family and sorted in lexicographic order
within the column family. For example, in a column family called
SysMonitorwith column qualifiers of
Priority, Cloud Bigtable stores the columns in this order:
- The intersection of a row and column can contain multiple timestamped cells. Each cell contains a unique, timestamped version of the data for that row and column.
- All operations are atomic at the row level. This means that an operation affects either an entire row or none of the row.
Ideally, both reads and writes should be distributed evenly across the row space of a table.
Cloud Bigtable tables are sparse. A column doesn't take up any space in a row that doesn't use the column.
A good schema results in excellent performance and scalability, and a poorly designed schema can lead to a poorly performing system. Every use case is different and requires its own design, but the following best practices apply to most use cases. Exceptions are noted.
Starting at the table level and working down to the row key level, the following sections describe the best practices for schema design:
All table elements, especially row keys, should be designed with planned read requests in mind. Check quotas and limits for recommended and hard size limits for all database elements.
Store datasets with similar schemas in the same table, rather than in separate tables.
In other database systems, you might choose to store data in multiple tables based on the subject and number of columns. In Cloud Bigtable, however, it is usually better to store all your data in one big table. You can assign a unique row key prefix to use for each dataset, so that Cloud Bigtable stores the related data in a contiguous range of rows that you can then query by row key prefix.
Cloud Bigtable has a limit of 1,000 tables per instance, but in most cases, you should have far fewer tables than that. Creating many small tables is a Cloud Bigtable anti-pattern for a few reasons:
- Sending requests to many different tables can increase backend connection overhead, resulting in increased tail latency.
- Having multiple tables of different sizes can disrupt the behind-the-scenes load balancing that makes Cloud Bigtable function well.
You might justifiably want a separate table for a completely different use case that requires a completely different schema, but you should not use separate tables for similar data. For example, you shouldn't create a new table because it's a new year or you have a new customer.
Put related columns in the same column family. When a row contains multiple values that are related to one another, it's a good practice to group the columns that contain those values in the same column family. Group data as closely as you can to avoid needing to design complex filters and so you get just the information that you need, but no more, in your most frequent read requests.
Create up to about 100 column families per table. Creating more than 100 column families may cause performance degradation.
Choose short but meaningful names for your column families. Names are included in the data that is transferred for each request.
Put columns that have different data retention needs in different column families. This is important if you want to limit storage costs. Garbage-collection policies are set at the column-family level, not at the column level. For example, if you only need to keep the most recent version of a particular piece of data, don't store it in a column family that is set to store 1,000 versions of something else. Otherwise you're paying to store 999 cells' worth of data you don't need.
Treat column qualifiers as data. Since you have to store a column qualifier for every column, you can save space by naming the column with a value. As an example, consider a table that stores data about friendships, in which each row represents a person and all their friendships. Each column qualifier can be the ID of a friend, and the value for that column in that row can be the social circle the friend is in. In this example, rows might look like this:
Contrast this with a schema for the same data that doesn't use column qualifiers as data:
The second schema design causes the table to grow much faster.
Create as many columns as you need in the table. Cloud Bigtable tables are sparse, and there is no space penalty for a column that is not used in a row. You can have millions of columns in a table, as long as no row exceeds 100 MB.
Avoid using too many columns in any single row. Even though a table can have millions of columns, a row should not. A few factors contribute to this best practice:
- It takes time for Cloud Bigtable to process each cell in a row.
- Each cell adds some overhead to the amount of data that's stored in your table and sent over the network. For example, if you're storing 1 KB (1,024 bytes) of data, it's much more space-efficient to store that data in a single cell, rather than spreading the data across 1,024 cells that each contain 1 byte.
If your dataset logically requires more columns per row than Cloud Bigtable can process efficiently, consider storing the data as a protobuf in a single column.
If you not are not using column qualifiers to store data and you want to reduce the amount of data that is transferred for each request, give column qualifiers short but meaningful names. The maximum size is 16 KB.
Don't store more than 100 MB of data in a single row. Rows that exceed this limit can result in reduced read performance.
Keep all information for an entity in a single row. For most uses cases, avoid storing data that you need to read atomically, or all at once, in more than one row to avoid inconsistencies. For example, if you update two rows in a table, it's possible that one row will be updated successfully and the other update will fail. Make sure your schema does not require more than one row to be updated at the same time in order for related data to be accurate. This ensures that if part of a write request fails or needs to be sent again, that piece of data is not temporarily incomplete.
Exception: If keeping an entity in a single row results in rows that are hundreds of MB in size, you should split the data across multiple rows.
Store related entities in adjacent rows, to make reads more efficient.
Don't store more than 10 MB of data in a single cell. Recall that a cell is the data stored for a given row and column with a unique timestamp, and that multiple cells can be stored at the intersection of that row and column. The number of cells retained in a column is governed by the garbage collection policy that you set for the column's column family.
Design your row key based on the queries you will use to retrieve the data. Well-designed row keys get the best performance out of Cloud Bigtable. The most efficient Cloud Bigtable queries retrieve data using one of the following:
- Row key
- Row key prefix
- Range of rows defined by starting and ending row keys
Other types of queries trigger a full table scan, which is much less efficient. By choosing the correct row key now, you can avoid a painful data migration process later.
Keep your row keys short. A row key must be 4 KB or less. Long row keys take up additional memory and storage and increase the time it takes to get responses from the Cloud Bigtable server.
Store multiple delimited values in each row key. Because the best way to query Cloud Bigtable efficiently is by row key, it's often useful to include multiple identifiers in your row key. When your row key includes multiple values, it's especially important to have a clear understanding of how you'll use your data.
Row key segments are usually separated by a delimiter, such as a colon, slash, or hash symbol. The first segment or set of contiguous segments is the row key prefix, and the last segment or set of contiguous segments is the row key suffix.
Well-planned row key prefixes let you take advantage of Cloud Bigtable's built-in sorting order to store related data in contiguous rows. Storing related data in contiguous rows lets you access related data as a range of rows, rather than running inefficient table scans.
If your data includes integers that you want to store or sort numerically, pad the integers with leading zeroes. Cloud Bigtable stores data lexicographically. For example, lexicographically, 3 > 20 but 20 > 03. Padding the 3 with a leading zero ensures the numbers are sorted numerically. This tactic is important for timestamps where range-based queries are desired.
It's important to create a row key that makes it possible to retrieve a well-defined range of rows. Otherwise, your query requires a table scan, which is much slower than retrieving specific rows.
For example, if your application tracks mobile device data, you can have a row key that consists of device type, device ID, and the day the data is recorded. Row keys for this data might look like this:
phone#4c410523#20200501 phone#4c410523#20200502 tablet#a0b81f74#20200501 tablet#a0b81f74#20200502
This row key design lets you retrieve data with a single request for:
- A device type
- A combination of device type and device ID
This row key design would not be optimal if you want to retrieve all data for a given day. Because the day is stored in the third segment, or the row key suffix, you cannot just request a range of rows based on the suffix or a middle segment of the row key. Instead, you have to send a read request with a filter that scans the entire table looking for the day value.
Use human-readable string values in your row keys whenever possible. This practice makes it easier to use the Key Visualizer tool to troubleshoot issues with Cloud Bigtable.
In many cases, you should design row keys that start with a common value and end with a granular value. For example, if your row key includes a continent, country, and city, you can create row keys that looks like the following so that they automatically sort first by values with lower cardinality:
asia#india#bangalore asia#india#mumbai asia#japan#okinawa asia#japan#sapporo southamerica#bolivia#cochabamba southamerica#bolivia#lapaz southamerica#chile#santiago southamerica#chile#temuco
Row keys to avoid
Some types of row keys can make it difficult to query your data, and some result in poor performance. This section describes some types of row keys that you should avoid using in Cloud Bigtable.
Row keys that start with a timestamp. This will cause sequential writes to be pushed onto a single node, creating a hotspot. If you put a timestamp in a row key, you need to precede it with a high-cardinality value like a user ID to avoid hotspotting.
Row keys that cause related data to not be grouped together. Avoid row keys that cause related data to be stored in non-contiguous row ranges, which are inefficient to read together.
Sequential numeric IDs. Suppose your system assigns a numeric ID to each of your application's users. You might be tempted to use the user's numeric ID as the row key for your table. However, because new users are more likely to be active users, this approach is likely to push most of your traffic to a small number of nodes.
A safer approach is to use a reversed version of the user's numeric ID, which spreads traffic more evenly across all of the nodes for your Cloud Bigtable table.
Frequently updated identifiers. Avoid using a single row key to identify a
value that must be updated very frequently. For example, if you store
memory-usage data for a number of devices once per second, do not use a
single row key for each device that is made up of the device ID and the metric
being stored, such as
4c410523#memusage, and update the row repeatedly. This
type of operation overloads the tablet that stores the frequently used row. It
can also cause a row to exceed its size limit, because a column's previous
values take up space until they are garbage-collected.
Instead, store each new reading in a new row. Using the memory usage example,
each row key can contain the device ID, the type of metric, and a timestamp, so
the row keys are similar to
strategy is efficient because in Cloud Bigtable, creating a new row
takes no more time than creating a new cell. In addition, this strategy enables
you to quickly read data from a specific date range by calculating the
appropriate start and end keys.
For values that change very frequently, such as a counter that is updated hundreds of times each minute, it's best to simply keep the data in memory, at the application layer, and write new rows to Cloud Bigtable periodically.
Hashed values. Hashing a row key removes your ability to take advantage of Cloud Bigtable's natural sorting order, making it impossible to store rows in a way that are optimal for querying. For the same reason, hashing values makes it challenging to use the Key Visualizer tool to troubleshoot issues with Cloud Bigtable. Use human-readable values instead of hashed values.
Values expressed as raw bytes rather than human-readable strings. Raw bytes are fine for column values, but for readability and troubleshooting, use string values in row keys.
Special use cases
You may have a unique dataset that requires special consideration when designing a schema to store it in Cloud Bigtable. This section describes some, but not all, different types of Cloud Bigtable data and some suggested tactics for storing it in the most optimal way.
Include a timestamp as part of your row key if you often need to retrieve data based on the time when it was recorded.
For example, your application might need to record performance-related data,
such as CPU and memory usage, once per second for a large number of machines.
Your row key for this data could combine an identifier for the machine with a
timestamp for the data (for example,
machine_4223421#1425330757685). Keep in
mind that row keys are sorted lexicographically.
Don't use a timestamp by itself or at the beginning of a row key, because this will cause sequential writes to be pushed onto a single node, creating a hotspot.
If you usually retrieve the most recent records first, you can use a reversed timestamp in the row key by subtracting the timestamp from your programming language's maximum value for long integers (in Java, java.lang.Long.MAX_VALUE). With a reversed timestamp, the records will be ordered from most recent to least recent.
For information specifically about working with time series data, see Schema design for time series data.
Row key prefixes provide a scalable solution for a "multi-tenancy" use case, a scenario in which you store similar data, using the same data model, on behalf of multiple clients. Using one table for all tenants is the most efficient way to store and access multi-tenant data.
For example, let's say you store and track purchase histories on behalf of many companies. You can use your unique ID for each company as a row key prefix. All data for a tenant is stored in contiguous rows in the same table, and you can query or filter using the row key prefix. Then, when a company is no longer your customer and you need to delete the purchase history data you were storing for the company, you can drop the range of rows that use that customer's row key prefix.
For example, if you are storing cell phone data for customers
examplepetstore, you can create row keys like the following. Then, if
altostrat is no longer your customer, you drop all rows with the row key
altostrat#phone#4c410523#20190501 altostrat#phone#4c410523#20190502 altostrat#tablet#a0b41f74#20190501 examplepetstore#phone#4c410523#20190502 examplepetstore#tablet#a6b81f79#20190501 examplepetstore#tablet#a0b81f79#20190502
In contrast, if you store data on behalf of each company in its own table, you can experience performance and scalability issues. You are also more likely to inadvertently reach Cloud Bigtable's limit of 1,000 tables per instance. After an instance reaches this limit, Cloud Bigtable prevents you from creating more tables in the instance.
Unless your use case demands it, avoid using personally identifiable information (PII) or user data in row keys or column family IDs. Row keys and column families are both data and metadata, and applications that use them as metadata, like encryption or logging, can inadvertently expose them to users who should not have access to private data.
Wide range of domain names
If you're storing data about entities that can be represented as domain names,
consider using a reverse domain name (for example,
com.company.product) as the
row key. Using a reverse domain name is an especially good idea if each row's
data tends to overlap with adjacent rows. In this case, Cloud Bigtable
can compress your data more efficiently.
In contrast, standard domain names that are not reversed can cause rows to be sorted in such a way that related data is not grouped together in one place, which can result in less efficient compression and less efficient reads.
This approach works best when your data is spread across many different reverse domain names.
To illustrate this point, consider the following domain names, automatically sorted in lexicographic order by Cloud Bigtable:
drive.google.com en.wikipedia.org maps.google.com
This is undesirable for the use case where you want to query all rows for the
google.com. In contrast, consider the same rows where the domain names have
com.google.drive com.google.maps en.wikipedia.org
In the second example, the related rows are automatically sorted in a way that makes it easy to retrieve them as a range of rows.
Few domain names
If you expect to store a lot of data for only one or a small number of domain names, consider other values for your row key. Otherwise, you might push writes to a single node in your cluster, resulting in hotspotting, or your rows might grow too large.
Changing or uncertain queries
If you don't always run the same queries on your data, or you are unsure what your queries will be, one option is to store all the data for a row in one column instead of multiple columns. With this approach, you use a format that makes it easy to extract the individual values later, such as the protocol buffer binary format or json.
The row key is still carefully designed to make sure you can retrieve the data you need, but each row typically has only one column that contains all the data for the row in a single protobuf.
Storing data as a protobuf message in one column instead of spreading the data into multiple columns has advantages and disadvantages. Advantages include the following:
- The data takes up less space, so it costs you less to store it.
- You maintain a certain amount of flexibility by not committing to column families and column qualifiers.
- Your reading application does not need to "know" what your table schema is.
Some disadvantages are the following:
- You have to deserialize the protobuf messages after they are read from Cloud Bigtable.
- You lose the option to query the data in protobuf messages using filters.
- You can't use BigQuery to run federated queries on fields within protobuf messages after reading them from Cloud Bigtable.
Learn how to design a schema for time-series data.
Review the steps involved in planning a schema.
Review the types of write requests you can send to Cloud Bigtable.