cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Aleksey Yeschenko (JIRA)" <>
Subject [jira] [Commented] (CASSANDRA-7447) New sstable format with support for columnar layout
Date Mon, 01 Sep 2014 10:29:23 GMT


Aleksey Yeschenko commented on CASSANDRA-7447:

It was my understanding that CASSANDRA-7743 was a temporary solution, there only until we
migrate all the date to the new format. And we absolutely do not want to keep two or more
formats indefinitely - getting rid of the old format *is* one of the goals.

As for the new format itself - yes, we do need it to support all the use cases (because we
want to ultimately have just one format, with support for all the things).

Question is how to deliver it incrementally while making sure that it will eventually (and
soon enough) support all the things.

I’d settle for listing all the features the initial format won’t support in the beginning,
and then detailed solutions for each of those features - but at that point we might as well
code it up. That, and I’m not sure we can guarantee we can even make such a list. There
are many cases, obscure ones, that nevertheless must be supported - empty values in primitive
types, mixed dynamic/static ‘thrift CFs’, non byteorder-comparable, incl. custom, types.
Having an upgrade path for all of those is a strict requirement that you can’t simply wish
away  (very unfortunately). It would be unacceptable to get close to 3.0 and only then realize
that oops, there is a use case that we can’t implement with the new format, gotta stick
with keeping both formats forever.

What would even be the logistics w/ a ‘feature’ not implemented by a new format? If the
format is set to NEW in the yaml, and someone writes an empty BB to an int column - what do
we do on flush? If someone adds a non-byteorder-comparable column, or a collection - at what
stage do we reject that?

That said, I think we are all in agreement that the current format can and should be replaced.
And I agree that we should go for the big win here. I just think that the real big win is
the one that would benefit all of our users (and “avoiding repeating clustering columns,
using a compact encoding for CQL column names, and use offset-encoding for timestamps” benefits
everyone - 1 and 3 even dense table users).

This way we can deliver something strictly better than the old format to everyone, without
the risk of getting stuck with two formats forever, and optimize for the special cases later.

> New sstable format with support for columnar layout
> ---------------------------------------------------
>                 Key: CASSANDRA-7447
>                 URL:
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Benedict
>            Assignee: Benedict
>              Labels: performance, storage
>             Fix For: 3.0
>         Attachments: ngcc-storage.odp
> h2. Storage Format Proposal
> C* has come a long way over the past few years, and unfortunately our storage format
hasn't kept pace with the data models we are now encouraging people to utilise. This ticket
proposes a collections of storage primitives that can be combined to serve these data models
more optimally.
> It would probably help to first state the data model at the most abstract level. We have
a fixed three-tier structure: We have the partition key, the clustering columns, and the data
columns. Each have their own characteristics and so require their own specialised treatment.
> I should note that these changes will necessarily be delivered in stages, and that we
will be making some assumptions about what the most useful features to support initially will
be. Any features not supported will require sticking with the old format until we extend support
to all C* functionality.
> h3. Partition Key
> * This really has two components: the partition, and the value. Although the partition
is primarily used to distribute across nodes, it can also be used to optimise lookups for
a given key within a node
> * Generally partitioning is by hash, and for the moment I want to focus this ticket on
the assumption that this is the case
> * Given this, it makes sense to optimise our storage format to permit O(1) searching
of a given partition. It may be possible to achieve this with little overhead based on the
fact we store the hashes in order and know they are approximately randomly distributed, as
this effectively forms an immutable contiguous split-ordered list (see Shalev/Shavit, or CASSANDRA-7282),
so we only need to store an amount of data based on how imperfectly distributed the hashes
are, or at worst a single value per block.
> * This should completely obviate the need for a separate key-cache, which will be relegated
to supporting the old storage format only
> h3. Primary Key / Clustering Columns
> * Given we have a hierarchical data model, I propose the use of a cache-oblivious trie
> * The main advantage of the trie is that it is extremely compact and _supports optimally
efficient merges with other tries_ so that we can support more efficient reads when multiple
sstables are touched
> * The trie will be preceded by a small amount of related data; the full partition key,
a timestamp epoch (for offset-encoding timestamps) and any other partition level optimisation
data, such as (potentially) a min/max timestamp to abort merges earlier
> * Initially I propose to limit the trie to byte-order comparable data types only (the
number of which we can expand through translations of the important types that are not currently)
> * Crucially the trie will also encapsulate any range tombstones, so that these are merged
early in the process and avoids re-iterating the same data
> * Results in true bidirectional streaming without having to read entire range into memory
> h3. Values
> There are generally two approaches to storing rows of data: columnar, or row-oriented.
The above two data structures can be combined with a value storage scheme that is based on
either. However, given the current model we have of reading large 64Kb blocks for any read,
I am inclined to focus on columnar support first, as this delivers order-of-magnitude benefits
to those users with the correct workload, while for most workloads our 64Kb blocks are large
enough to store row-oriented data in a column-oriented fashion without any performance degradation
(I'm happy to consign very large row support to phase 2). 
> Since we will most likely target both behaviours eventually, I am currently inclined
to suggest that static columns, sets and maps be targeted for a row-oriented release, as they
don't naturally fit in a columnar layout without secondary heap-blocks. This may be easier
than delivering heap-blocks also, as it keeps both implementations relatively clean. This
is certainly open to debate, and I have no doubt there will be conflicting opinions here.
> Focusing on our columnar layout, the goals are:
> * Support sparse and dense column storage
> * Efficient compression of tombstones, timestamps, ttls, etc. into near-zero space based
on offset/delta/bitmap encoding
> * Normalisation of column names once per file
> * Per-file block-layout index, defining how each block's data is encoded, so we can index
directly within a block for dense fields (permitting more efficient page cache utilisation)
> * Configurable grouping of fields per block, so that we can get closer to row-oriented
or column-oriented performance, based on user goals
> I have attached my slides from the ngcc for reference.

This message was sent by Atlassian JIRA

View raw message