Was this page helpful?
You're viewing documentation for a previous version of ScyllaDB Enterprise. Switch to the latest stable version.
Scylla supports the DynamoDB API (this feature is codenamed “Alternator”). Our goal is to support any application written for Amazon DynamoDB. Nevertheless, there are a few differences between DynamoDB and Scylla, and and a few DynamoDB features that have not yet been implemented in Scylla. The purpose of this document is to inform users of these differences.
The most obvious difference between DynamoDB and Scylla is that while DynamoDB is a shared cloud service, Scylla is a dedicated service running on your private cluster. Whereas DynamoDB allows you to “provision” the number of requests per second you’ll need - or at an extra cost not even provision that - Scylla requires you to provision your cluster. You need to reason about the number and size of your nodes - not the throughput.
When creating a table, the BillingMode and ProvisionedThroughput options are ignored by Scylla.
DynamoDB applications specify a single “endpoint” address, e.g.,
dynamodb.us-east-1.amazonaws.com. Amazon’s cloud service distributes
request for this single URL to many different backend nodes. Such a
load-balancing setup is not included inside Alternator. You should either
set one up, or configure the client library to do the load balancing itself.
Instructions for doing this can be found in:
Scylla was designed to optimize the performance of pure write operations - writes which do not need to read the the previous value of the item. In CQL, writes which do need the previous value of the item must explicitly use the slower LWT (”LightWeight Transaction”) feature to be correctly isolated from each other. It is not allowed to mix LWT and non-LWT writes to the same item.
In contrast, in the DynamoDB API every write operation may need the previous value of the item. So without making further assumptions, Alternator would need to use the slower LWT for all writes - to correctly isolate concurrent writes. However, if Alternator is told that a certain workload does not have any read-modify-write operations, it can do all writes with the faster non-LWT write. Furthermore, if Alternator is told that a certain workload does have do both write-only and read-modify-write, but to different items, it could use LWT only for the read-modify-write operations.
Therefore, Alternator must be explicitly configured to tell it which of the
above assumptions it may make on the write workload. This configuration is
mandatory, and described in the “Write isolation policies” section of
alternator.md. One of the options,
always_use_lwt, is always safe, but the
other options result in significantly better write performance and should be
considered when the workload involves pure writes (e.g., ingestion of new
data) or if pure writes and read-modify-writes go to distinct items.
When a DynamoDB application writes twice to the same item, it expects “last-
write-wins” behavior: The later write should overwrite the earlier write.
When writes use LWT (the
always_use_lwt policy described above), this is
indeed guaranteed. However, for other write isolation policies, Scylla
does not guarantee that writes are not reordered. In some sense, the “last”
write does still win, but the meaning of which write is “last” is different
from what most users expect:
In this case (write isolation policy is not
always_use_lwt), each write
request gets a timestamp which is the current time on the server which
received this request. If two write requests arrive at different Alternator
nodes, and if the local clocks on these two nodes are not accurately
synchronized, then the two timestamps generated independently on the two
nodes may have the opposite order as intended - the earlier write may get
a higher timestamp - and this will be the “last write” that wins.
To avoid or mitigate this write reordering issue, users may consider one or more of the following:
Use NTP to keep the clocks on the different Scylla nodes synchronized. If the delay between the two writes is longer than NTP’s accuracy, they will not be reordered.
If an application wants to ensure that two specific writes are not reordered, it should send both requests to the same Scylla node. Care should be taken when using a load balancer - which might redirect two requests to two different nodes.
Consider using the
always_use_lwt write isolation policy.
It is slower, but has better guarantees.
Another guarantee that that
always_use_lwt can make and other write
isolation modes do not is that writes to the same item are serialized:
Even if the two write are sent at exactly the same time to two different
nodes, the result will appear as if one write happended first, and then
the other. But in other modes (with non-LWT writes), two writes can get
exactly the same microsecond-resolution timestamp, the the result may be
a mixture of both writes - some attributes from one and some from the
other - instead of being just one or the other.
Alternator implements the same signature protocol as DynamoDB and the rest of AWS. Clients use, as usual, an access key ID and a secret access key to prove their identity and the authenticity of their request. Alternator can then validate the authenticity and authorization of each request using a known list of authorized key pairs.
In the current implementation, the user stores the list of allowed key pairs
system_auth.roles table: The access key ID is the
role column, and
the secret key is the
salted_hash, i.e., the secret key can be found by
SELECT salted_hash from system_auth.roles WHERE role = ID;.
By default, authorization is not enforced at all. It can be turned on
by providing an entry in Scylla configuration:
Although Alternator implements DynamoDB’s authentication, including the possibility of listing multiple allowed key pairs, there is currently no implementation of access control. All authenticated key pairs currently get full access to the entire database. This is in contrast with DynamoDB which supports fine-grain access controls via “IAM policies” - which give each authenticated user access to a different subset of the tables, different allowed operations, and even different permissions for individual items. All of this is not yet implemented in Alternator.
Scylla has an advanced and extensive monitoring framework for inspecting and graphing hundreds of different metrics of Scylla’s usage and performance. Scylla’s monitoring stack, based on Grafana and Prometheus, is described in https://docs.scylladb.com/operating-scylla/monitoring/. This monitoring stack is different from DynamoDB’s offering - but Scylla’s is significantly more powerful and gives the user better insights on the internals of the database and its performance.
Like in DynamoDB, Alternator items which are set to expire at a certain time will not disappear exactly at that time, but only after some delay. DynamoDB guarantees that the expiration delay will be less than 48 hours (though for small tables the delay is often much shorter).
In Alternator, the expiration delay is configurable - it can it can be be set
--alternator-ttl-period-in-seconds configuration option.
The default is 24 hours.
In ScyllaDB versions 5.0 and earlier, DynamoDB’s TTL feature
is experimental and needs to be enabled explicitly with the
--experimental-features=alternator-ttl configuration option.
See Enabling Experimental Features for details.
One thing that the experimental implementation is missing is that expiration events appear in the Streams API as normal deletions - without the distinctive marker on deletions which are really expirations. See https://github.com/scylladb/scylla/issues/5060.
Note This feature is experimental in versions earlier than ScyllaDB Open Source 5.2 and ScyllaDB Enterprise 2022.2.
Some DynamoDB API features are supported by Alternator, but considered experimental in this release. An experimental feature in Scylla is a feature whose functionality is complete, or mostly complete, but it is not as thoroughly tested or optimized as regular features. Also, an experimental feature’s implementation is still subject to change and upgrades may not be possible if such a feature is used. For these reasons, experimental features are not recommended for mission-critical uses, and they need to be individually enabled with the “–experimental-features” configuration option.
In this release, the following DynamoDB API features are considered experimental:
The DynamoDB Streams API for capturing change is supported, but still
considered experimental so needs to be enabled explicitly with the
--experimental-features=alternator-streams configuration option.
Alternator streams also differ in some respects from DynamoDB Streams:
The number of separate “shards” in Alternator’s streams is significantly larger than is typical on DynamoDB.
While in DynamoDB data usually appears in the stream less than a second after it was written, in Alternator Streams there is currently a 10 second delay by default. https://github.com/scylladb/scylla/issues/6929
Some events are represented differently in Alternator Streams. For example, a single PutItem is represented by a REMOVE + MODIFY event, instead of just a single MODIFY or INSERT. https://github.com/scylladb/scylla/issues/6930 https://github.com/scylladb/scylla/issues/6918
In general, every DynamoDB API feature available in Amazon DynamoDB should behave the same in Alternator. However, there are a few features which we have not implemented yet. Unimplemented features return an error when used, so they should be easy to detect. Here is a list of these unimplemented features:
Currently in Alternator, a GSI (Global Secondary Index) can only be added to a table at table creation time. Unlike DynamoDB which also allows adding a GSI (but not an LSI) to an existing table using an UpdateTable operation. https://github.com/scylladb/scylla/issues/5022
GSI (Global Secondary Index) and LSI (Local Secondary Index) may be configured to project only a subset of the base-table attributes to the index. This option is not yet respected by Alternator - all attributes are projected. This wastes some disk space when it is not needed. https://github.com/scylladb/scylla/issues/5036
DynamoDB’s new multi-item transaction feature (TransactWriteItems, TransactGetItems) is not supported. Note that the older single-item conditional updates feature are fully supported. https://github.com/scylladb/scylla/issues/5064
Alternator does not yet support the DynamoDB API calls that control which table is available in which data center (DC): CreateGlobalTable, UpdateGlobalTable, DescribeGlobalTable, ListGlobalTables, UpdateGlobalTableSettings, DescribeGlobalTableSettings, and UpdateTable. Currently, all Alternator tables are created as global tables and can be accessed from all the DCs existing at the time of the table’s creation. If a DC is added after a table is created, the table won’t be visible from the new DC and changing that requires a CQL “ALTER TABLE” statement to modify the table’s replication strategy. https://github.com/scylladb/scylla/issues/5062
Recently DynamoDB added support, in addition to the DynamoDB Streams API, also for the similar Kinesis Streams. Alternator doesn’t support this yet, and the related operations DescribeKinesisStreamingDestination, DisableKinesisStreamingDestination, and EnableKinesisStreamingDestination. https://github.com/scylladb/scylla/issues/8786
The on-demand backup APIs are not supported: CreateBackup, DescribeBackup, DeleteBackup, ListBackups, RestoreTableFromBackup. For now, users can use Scylla’s existing backup solutions such as snapshots or Scylla Manager. https://github.com/scylladb/scylla/issues/5063
Continuous backup (the ability to restore any point in time) is also not supported: UpdateContinuousBackups, DescribeContinuousBackups, RestoreTableToPointInTime
DynamoDB’s encryption-at-rest settings are not supported. The Encryption- at-rest feature is available in Scylla Enterprise, but needs to be enabled and configured separately, not through the DynamoDB API.
No support for throughput accounting or capping. As mentioned above, the BillingMode option is ignored by Alternator, and if a provisioned throughput is specified, it is ignored. Requests which are asked to return the amount of provisioned throughput used by the request do not return it in Alternator. https://github.com/scylladb/scylla/issues/5068
DAX (DynamoDB Accelerator), an in-memory cache for DynamoDB, is not available in for Alternator. Anyway, it should not be necessary - Scylla’s internal cache is already rather advanced and there is no need to place another cache in front of the it. We wrote more about this here: https://www.scylladb.com/2017/07/31/database-caches-not-good/
The DescribeTable is missing information about creation data and size estimates, and also part of the information about indexes enabled on the table. https://github.com/scylladb/scylla/issues/5013 https://github.com/scylladb/scylla/issues/5026 https://github.com/scylladb/scylla/issues/7550 <https://github.com/scylladb/scylla/issues/7551 >
The recently-added PartiQL syntax (SQL-like SELECT/UPDATE/INSERT/DELETE expressions) and the new operations ExecuteStatement, BatchExecuteStatement and ExecuteTransaction is not yet supported. A user that is interested in an SQL-like syntax can consider using Scylla’s CQL protocol instead. https://github.com/scylladb/scylla/issues/8787
As mentioned above, Alternator has its own powerful monitoring framework, which is different from AWS’s. In particular, the operations DescribeContributorInsights, ListContributorInsights and UpdateContributorInsights that configure Amazon’s “CloudWatch Contributor Insights” are not yet supported. Scylla has different ways to retrieve the same information, such as which items were accessed most often. https://github.com/scylladb/scylla/issues/8788
Alternator does not support the new DynamoDB feature “export to S3”, and its operations DescribeExport, ExportTableToPointInTime, ListExports. https://github.com/scylladb/scylla/issues/8789
Was this page helpful?