Proton is Vespa's search core. Proton maintains disk and memory structures for documents. As the data is dynamic, these structures are periodically optimized by maintenance jobs and the resource footprint of these background jobs are primarily controlled by the concurrency setting.
The content node has a bucket management system
which sends requests to a set of document databases,
which each consists of three sub-databases
not ready and
When the node starts up it first needs to get an overview of what documents and buckets it has. Once it knows this, it is in initializing mode, able to handle load, but distributors do not yet know bucket metadata for all the buckets, and thus can't know whether buckets are consistent with copies on other nodes. Once metadata for all buckets are known, the content nodes transitions from initializing to up state. As the distributors wants quick access to bucket metadata, it keeps an in-memory bucket database to efficiently serve these requests.
It implements elasticity support in terms of the SPI. Operations are ordered according to priority, and only one operation per bucket can be in-flight at a time. Below bucket management is the persistence engine, which implements the SPI in terms of Vespa search. The persistence engine reads the document type from the document id, and dispatches requests to the correct document database.
Each document database is responsible for a single document type. It has a component called FeedHandler which takes care of incoming documents, updates, and remove requests. All requests are first written to a transaction log, then handed to the appropriate sub-database, based on the request type.
There are three types of sub-databases, each with its own document meta store and document store. The document meta store holds a map from the document id to a local id. This local id is used to address the document in the document store. The document meta store also maintains information on the state of the buckets that are present in the sub-database.
The sub-databases are maintained by the index maintainer. The document distribution changes as the system is resized. When the number of nodes in the system changes, the index maintainer will move documents between the Ready and Not Ready sub-databases to reflect the new distribution. When an entry in the Removed sub-database gets old it is purged. The sub-databases are:
|Not Ready||Holds the redundant documents that are not searchable, i.e. the not ready documents. Documents that are not ready are only stored, not indexed. It takes some processing to move from this state to the ready state.|
Maintains an index of all ready documents and keeps them searchable.
One of the ready copies is active while the rest are not active:
|Removed||Keeps track of documents that have been removed. The id and timestamp for each document is kept. This information is used when buckets from two nodes are merged. If the removed document exists on another node but with a different timestamp, the most recent entry prevails.|
Content nodes have a transaction log to persist mutating operations. The transaction log persists operations by file append. Having a transaction log simplifies proton's in-memory index structures and enables steady-state high performance, read more below.
All operations are written and synched to the transaction log. This is sequential (not random) IO but might impact overall feed performance if running on NAS attached storage where the synch operation has a much higher cost than on local attached storage (e.g. SSD). See sync-transactionlog.
Default, proton will flush components like attribute vectors and memory index on shutdown, for quicker startup after scheduled restarts.
Proton stores position information in text indices by default, for proximity relevance -
All the occurrences of a term is stored in the posting list, with its position.
This provides superior ranking features,
but is somewhat more expensive than just storing a single occurrence per document.
For most applications it is the correct tradeoff,
since most of the cost is usually elsewhere and relevance is valuable.
Applications that only need occurrence information for filtering
can use rank: filter
to optimize query performance, using only
The memory index has a dictionary per index field. This contains all unique words in that field with mapping to posting lists with position information. The position information is used during ranking, see nativeRank for details on how a text match score is calculated.
The disk index stores the content of each index field in separate folders. Each folder contains:
boolocc-files are empty if number of occurrences is small, like in the example above.
Documents are stored in the document store. Put, update and remove operations are persisted in the transaction log before updating the document in the document store. The operation is ack'ed to the client and the result of the operation is immediately seen in search results.
Files in the document store are written sequentially, and occur in pairs - example:
-rw-r--r-- 1 owner users 4133380096 Aug 10 13:36 1467957947689211000.dat -rw-r--r-- 1 owner users 71192112 Aug 10 13:36 1467957947689211000.idx
The maximum size: (in bytes) per .dat file on disk can be set using the following:
<content id="mycluster" version="1.0"> <engine> <proton> <tuning> <searchnode> <summary> <store> <logstore> <maxfilesize>8000000000</maxfilesize>Notes:
Document store compaction, defragments and sort document store files. It removes stale versions of documents (i.e. old version of updated documents). It is triggered when the disk bloat of the document store is larger than the total disk usage of the document store times diskbloatfactor. Refer to summary tuning for details.
Defragmentation status is best observed by tracking max_bucket_spread over time, a sawtooth pattern is normal for corpora that change over time. The document_store_compact metric tracks when proton is running compaction jobs. Compaction settings can be set too tight, in that case, the metric is always, or close to, 1.
When benchmarking, it is important to set the correct compaction settings, and also make sure that proton has compacted files since (can take hours), and is not actively compacting (document_store_compact should be 0 most of the time).
As documents are clustered within the .dat file, proton optimizes reads by reading larger chunks when accessing documents. When visiting, documents are read in bucket order. This is the same order as the defragmentation jobs uses.
The first document read in a visit operation for a bucket will read a chunk from the .dat file into memory. Subsequent document accesses are served be a memory lookup only. The chunk size is configured by maxsize:
<engine> <proton> <tuning> <searchnode> <summary> <store> <logstore> <chunk> <maxsize>16384</maxsize> </chunk> </logstore>
There can be 2^22=4M chunks. This sets a minimum chunk size based on maxfilesize - e.g. an 8G file can have minimum 2k chunk size. Finally, bucket size is configured by setting bucket-splitting:
<content id="imagepersonal" version="1.0"> <tuning> <bucket-splitting max-documents="1024"/>
The following are the relevant sizing units:
The document store has a mapping in memory from local ID (LID) to position in a document store file (.dat). Part of this mapping is persisted in the .idx-file paired to the .dat file. The memory used by the document store is linear with number of documents and updates to these.
The metric content.proton.documentdb.ready.document_store.memory_usage.allocated_bytes gives the size in memory - use the metric API to find it. A rule of thumb is 12 bytes per document.
There is only one instance of each job at a time - e.g. attributes are flushed in sequence. When a job is running, its metric is set to 1 - otherwise 0. Use this to correlate observed performance with job runs - see Run metric.
Flush an attribute vector from memory to disk, based on configuration in the flushstrategy. This controls memory usage and query performance. This also makes proton starts quicker - see flush on shutdown.
|CPU||Little - one thread flushes to disk|
|Memory||Little - some temporary use|
|Disk||A new file is written too, so 2x the size of an attribute on disk.|
|memory index flush||
Flush a memory index to disk, then trigger disk index fusion. The goal is to shrink memory usage by adding to the disk-backed indices. Performance characteristics for this flush is similar to indexing. Note: A high feed rate can cause multiple smaller flushed indices, like $VESPA_HOME/var/db/vespa/search/cluster.name/n1/documents/doc/0.ready/index/index.flush.102 - see the high index number. Multiple smaller indices is a symptom of too small memory indices compared to feed rate - to fix, increase flushstrategy > native > component > maxmemorygain.
|CPU||Little - one thread indexes to disk|
|Disk||Creates a new disk index, size of the memory index.|
|disk index fusion||
Merge the primary disk index with smaller indices generated by memory index flush - triggered by the memory index flush.
|CPU||Multiple threads merge indices, configured as a function of feeding concurrency - refer to this for details|
|Disk||Creates a new index while serving from the current: 2x temporary disk usage for the given index.|
|document store flush||
Flushes the document store.
|document store compaction|
|CPU||Little - one thread reads one files, sorts and writes a new file|
|Memory||Holds a document summary store file in memory plus memory for sorting the file. Note: This is important on hosts with little memory! Reduce maxfilesize to increase number of files and use less temporary memory for compaction.|
|Disk||A new file is written while the current is serving, max temporary usage is 2x.|
Triggered by nodes going up/down, refer to content cluster elasticity and searchable-copies. It causes documents to be indexed or de-indexed, similar to feeding. This moves documents in or out of ready/active sub-databases.
|CPU||CPU similar to feeding. Consumes capacity from the index write thread, so has feeding impact|
|Memory||As feeding - the memory index will grow|
As bucket move, however moves documents within a sub-database. This is often triggered when a cluster grows with more nodes, documents are redistributed to new nodes and each node has fewer documents - a LIDspace compaction is hence triggered. This in-place defragments the document meta store. Resources are freed on a subsequent attribute flush.
|CPU||like feeding - add and delete doc|
|removed documents pruning||
Prunes the deleted documents sub-database which keeps IDs for deleted documents. Default runs once per hour.
When the content node receives a get request, it scans through all the document databases, and for each one it checks all three sub-databases. Once the document is found, the scan is stopped and the document returned. If the document is found in a Ready sub-database, the document retriever will apply any changes that is stored in the attributes before returning the document.
A visit request creates an iterator over each candidate bucket. This iterator will retrieve matching documents from all sub-databases of all document databases. As for get, attributes values are applied to document fields in the Ready sub-database.
Queries has a separate pathway through the system. It does not use the distributor, nor has it anything to do with the SPI. It is orthogonal to the elasticity set up by the storage and retrieval described above. How queries move through the system:
A query enters the system through the QR-server (query rewrite server) in the Vespa Container. The QR-server issues one query per document type to the search nodes:
The Container knows all the document types and rewrites queries as a collection of queries, one for each type. Queries may have a restrict parameter, in which case the container will send the query only to the specified document types.
It sends the query to content nodes and collect partial results. It pings all content nodes every second to know whether they are alive, and keeps open TCP connections to each one. If a node goes down, the elastic system will make the documents available on other nodes.
|Content node matching||
The match engine receives queries and routes them to the right document database based on the document type. The query is passed to the Ready sub-database, where the searchable documents are. Based on information stored in the document meta store, the query is augmented with a blocklist that ensures only active documents are matched.
This section describes the custom extensions of the proton custom component state API.
Component status can be found by HTTP GET at
This gives an overview of the relevant search node components and their internal state.
Note that this is not a stable API, and it will expand and change between releases.