Data system partition design - partition and secondary index

Current zoning schemes rely on KV Data model .KV Simple model , It's all through K Access records , Naturally, it can be based on K Determine the partition , And route the read-write request to the responsible K The partition .

But if secondary index is involved , It's complicated . Secondary indexes usually do not uniquely identify a record , It's a query that speeds up specific values , Such as querying users JavaEdge All operations , Find words that contain java All blogs, etc .

many KV Storage （ Such as HBase） In order to reduce the implementation complexity, we give up the secondary index , But some （ Such as Riak） They have started to be supported , The secondary index is also Solr and ES Wait for the root of search server .

The main challenge of secondary indexes is that they cannot be mapped neatly to partitions . There are two schemes that support partitioning secondary indexes ：

• Document based partitioning （document-based）
• Based on keywords （term-based） The partition

3.1 Partition based on the secondary index of the document

Used car sales network （ Pictured -4）. Each list has a unique document ID, This is right for DB partition , Such as zoning 0 Medium ID 0~499, Partition 1 Medium ID 500~999.

User search , It can be filtered by color and manufacturer , So we need to set secondary index in color and manufacturer （ In the document DB These are fields （field）, Relationship DB These are columns （column））. Whenever a red car is added to DB,DB Partitions are automatically added to index entries color:red Documents ID list .

In this indexing method , Each partition is completely independent , Each maintains its own secondary index , And only responsible for the documents in your own partition , And don't care about the data of other partitions . Whenever you need to write DB（ add to , Delete or update documents ）, Just deal with the document containing the object you are writing ID The partition . therefore , Document partition index is also called local index , Not the global index .

But pay attention when reading ： Except for documents ID Special treatment , Otherwise, it is unlikely to put all cars of a specific color or brand in the same zone . chart -4 in , The red car appears in the partition 0、1. therefore , If you search the red car , You need to send the query to all partitions , Then merge all the returned results .

This query partition DB The method of is sometimes called decentralization / Gather （scatter/gather）, Obviously, this kind of secondary index query is expensive . Even parallel query partitions , Dispersed / Aggregation also tends to cause significant amplification of tail read latency . But it is still widely used ：MongoDB,Cassandra,ES Up to the secondary index based on document partition . Most of the DB The supplier suggests that users build a suitable partition scheme by themselves , Try to satisfy the secondary index query by a single partition , But this is not always feasible , Especially when multiple secondary indexes are used in queries （ For example, you need to press the color at the same time 、 The manufacturer inquires with two conditions ）.

3.2 Based on entries (Term) Secondary index partition of

A global index can be built for all data , Instead of each partition maintaining its own secondary index （ Local index ）. To avoid becoming a bottleneck , You cannot store a global index on a node , Otherwise, the purpose of setting partition balance will be destroyed . therefore , The global index must also be partitioned , But it can be used with K Different partition strategies .

Pictured -5, The red car of all data partitions is included in the index color:red, The index itself is partitioned , If you follow a To r The starting color is in the partition 0,s To z Partition 1. Allied , The index of automobile manufacturers is also partitioned （ The boundaries of the two partitions are f、 h）.

This index is called Entry partition （term-partitioned）, Take the keyword to be searched itself as the index . Such as color ：color:red. key word （Term） This name comes from full-text index （ A special secondary index ）,term Refers to all word sets that appear in the document .

Directly through key word Divide the index globally by itself , Or to its hash. Partitioning according to the keywords themselves is useful for range scanning （ For example, for properties of numeric classes ,e.g. Car quotation ）, And for keywords hash Partitions can be more evenly divided .

Global entry partition V.S Document partition index

• It makes reading more efficient , That is, there is no need to disperse / The collection performs a query on all partitions . contrary , The client only needs to send a read request to the partition containing the entry
• Disadvantages of global indexing , Writing is slow and complex , Because the update of a single document is , May affect multiple secondary indexes , The partitions of the secondary index may be located in different partitions or different nodes ,

Ideally , The index should be kept up to date , That is, every data written should be immediately reflected in the latest index . But partition entries , This requires distributed transactions across partitions , Writing speed will be greatly affected , So there are DB Synchronous updating of secondary indexes is not supported .

In practice , Updates to the global secondary index are asynchronous （ That is, if you read the index immediately after writing , Then the update may not be reflected in the index ）.