Design Concepts

Transactions: FAQ

Dgraph supports distributed ACID transactions through snapshot isolation.

Can we do pre-writes only on leaders?

Seems like a good idea, but has bad implications. If we only do a prewrite in-memory, only on leader, then this prewrite wouldn’t make it to the Raft log, or disk; but would be considered successful.

Then zero could mark the transaction as committed; but this leader could go down, or leadership could change. In such a case, we’d end up losing the transaction altogether despite it having been considered committed.

Therefore, pre-writes do have to make it to disk. And if so, better to propose them in a Raft group.

Consistency Model

[Update on Jan 15, 2020]

• Dgraph supports MVCC, Read Snapshots and Distributed ACID transactions.
• The transactions are cluster-wide (not key-only, or any other “crippled” version of them).
• Transactions are lockless. They don’t block/wait on seeing pending writes by uncommitted transactions. Zero would choose to commit or abort them depending on conflicts.
• Transactions are based on Snapshot Isolation (not Serializable Snapshot Isolation), because conflicts are determined by writes (not reads).
• Dgraph hands out monotonically increasing timestamps (for transactions). Ergo, if any transaction Tx1 commits before Tx2 starts, then Ts_commit(Tx1) < Ts_start(Tx2).
• Any commit at Tc are guaranteed to be seen by a read at timestamp Tr by any client, if Tr > Tc.
• All reads are snapshots across the entire cluster, seeing all previously committed transactions in full.

---

Consistency Models

[Last updated: Mar 2018. This is outdated and is not how we do things anymore] Basing it on this article by aphyr.

• Sequential Consistency: Different users would see updates at different times, but each user would see operations in order.

Dgraph has a client-side sequencing mode, which provides sequential consistency.

Here, let’s replace a “user” with a “client” (or a single process). In Dgraph, each client maintains a linearizable read map (linread map). Dgraph’s data set is sharded into many “groups”. Each group is a Raft group, where every write is done via a “proposal.” You can think of a transaction in Dgraph, to consist of many group proposals.

The leader in Raft group always has the most recent proposal, while replicas could be behind the leader in varying degrees. You can determine this by just looking at the latest applied proposal ID. A leader’s proposal ID would be greater than or equal to some replicas’ applied proposal ID.

linread map stores a group -> max proposal ID seen, per client. If a client’s last read had seen updates corresponding to proposal ID X, then linread map would store X for that group. The client would then use the linread map to inform future reads to ensure that the server servicing the request, has proposals >= X applied before servicing the read. Thus, all future reads, irrespective of which replica it might hit, would see updates for proposals >= X. Also, the linread map is updated continuously with max seen proposal IDs across all groups as reads and writes are done across transactions (within that client).

In short, this map ensures that updates made by the client, or seen by the client, would never be unseen; in fact, they would be visible in a sequential order. There might be jumps though, for e.g., if a value X → Y → Z, the client might see X, then Z (and not see Y at all).

• Linearizability: Each op takes effect atomically at some point between invocation and completion. Once op is complete, it would be visible to all.

Dgraph supports server-side sequencing of updates, which provides linearizability. Unlike sequential consistency which provides sequencing per client, this provide sequencing across all clients. This is necessary to make transactions work across clients. Thus, once a transaction is committed, it would be visible to all future readers, irrespective of client boundaries.

• Causal consistency: Dgraph does not have a concept of dependencies among transactions. So, does NOT order based on dependencies.
• Serializable consistency: Dgraph does NOT allow arbitrary reordering of transactions, but does provide a linear order per key.

Concepts

Edges

Typical data format is RDF N-Quad which is:

• Subject, Predicate, Object, Label, aka
• Entity, Attribute, Other Entity / Value, Label

Both the terminologies get used interchangeably in our code. Dgraph considers edges to be directional, i.e. from Subject -> Object. This is the direction that the queries would be run.

Internally, the RDF N-Quad gets parsed into this format.

type DirectedEdge struct {
  Entity      uint64
  Attr        string
  Value       []byte
  ValueType   uint32
  ValueId     uint64
  Label       string
  Lang 	      string
  Op          DirectedEdge_Op // Set or Delete
  Facets      []*facetsp.Facet
}

Note that irrespective of the input, both Entity and Object/ValueId get converted in UID format.

Posting List

Conceptually, a posting list contains all the DirectedEdges corresponding to an Attribute, in the following format:

Attribute: Entity -> sorted list of ValueId // Everything in uint64 representation.

So, for, e.g., if we’re storing a list of friends, such as:

Entity	Attribute	ValueId
Me	friend	person0
Me	friend	person1
Me	friend	person2
Me	friend	person3

Then a posting list friend would be generated. Seeking for Me in this PL would produce a list of friends, namely [person0, person1, person2, person3].

The big advantage of having such a structure is that we have all the data to do one join in one Posting List. This means, one RPC to the machine serving that Posting List would result in a join, without any further network calls, reducing joins to lookups.

Implementation wise, a Posting List is a list of Postings. This is how they look in Protocol Buffers format.

message Posting {
  fixed64 uid = 1;
  bytes value = 2;
  enum ValType {
    DEFAULT = 0;
    BINARY = 1;
    INT = 2; // We treat it as int64.
    FLOAT = 3;
    BOOL = 4;
    DATE = 5;
    DATETIME = 6;
    GEO = 7;
    UID = 8;
    PASSWORD = 9;
    STRING = 10;

  }
  ValType val_type = 3;
  enum PostingType {
    REF=0;          // UID
    VALUE=1;        // simple, plain value
    VALUE_LANG=2;   // value with specified language
        // VALUE_TIMESERIES=3; // value from timeseries, with specified timestamp
  }
  PostingType posting_type = 4;
  bytes metadata = 5; // for VALUE_LANG: Language, for VALUE_TIMESERIES: timestamp, etc..
  string label = 6;
  uint64 commit = 7;  // More inclination towards smaller values.
  repeated facetsp.Facet facets = 8;

  // TODO: op is only used temporarily. See if we can remove it from here.
  uint32 op = 12;
}

message PostingList {
  repeated Posting postings = 1;
  bytes checksum = 2;
  uint64 commit = 3; // More inclination towards smaller values.
}

There is typically more than one Posting in a PostingList.

The RDF Label is stored as label in each posting.

Badger

PostingLists are served via Badger, given the latter provides enough knobs to decide how much data should be served out of memory, SSD or disk. Also, it supports bloom filters on keys, which makes random lookups efficient.

To allow Badger full access to memory to optimize for caches, we’ll have one Badger instance per machine. Each instance would contain all the posting lists served by the machine.

Posting Lists get stored in Badger, in a key-value format, like so:

(Predicate, Subject) --> PostingList

Group

Every Alpha server belongs to a particular group, and each group is responsible for serving a particular set of predicates. Multiple servers in a single group replicate the same data to achieve high availability and redundancy of data.

Predicates are automatically assigned to each group based on which group first receives the predicate. By default periodically predicates can be moved around to different groups upon heuristics to evenly distribute the data across the cluster. Predicates can also be moved manually if desired.

In a future version, if a group gets too big, it could be split further. In this case, a single Predicate essentially gets divided across two groups.

  Original Group:
            (Predicate, Sa..z)
  After split:
  Group 1:  (Predicate, Sa..i)
  Group 2:  (Predicate, Sj..z)

Note that keys are sorted in BadgerDB. So, the group split would be done in a way to maintain that sorting order, i.e. it would be split in a way where the lexicographically earlier subjects would be in one group, and the later in the second.

Replication and Server Failure

Each group should typically be served by atleast 3 servers, if available. In the case of a machine failure, other servers serving the same group can still handle the load in that case.

New Server and Discovery

Dgraph cluster can detect new machines allocated to the cluster, establish connections, and transfer a subset of existing predicates to it based on the groups served by the new machine.

Write Ahead Logs

Every mutation upon hitting the database doesn’t immediately make it on disk via BadgerDB. We avoid re-generating the posting list too often, because all the postings need to be kept sorted, and it’s expensive. Instead, every mutation gets logged and synced to disk via append only log files called write-ahead logs. So, any acknowledged writes would always be on disk. This allows us to recover from a system crash, by replaying all the mutations since the last write to Posting List.

Mutations

In addition to being written to Write Ahead Logs, a mutation also gets stored in memory as an overlay over immutable Posting list in a mutation layer. This mutation layer allows us to iterate over Postings as though they’re sorted, without requiring re-creating the posting list.

When a posting list has mutations in memory, it’s considered a dirty posting list. Periodically, we re-generate the immutable version, and write to BadgerDB. Note that the writes to BadgerDB are asynchronous, which means they don’t get flushed out to disk immediately, but that wouldn’t lead to data loss on a machine crash. When Posting lists are initialized, write-ahead logs get referred, and any missing writes get applied.

Every time we regenerate a posting list, we also write the max commit log timestamp that was included – this helps us figure out how long back to seek in write-ahead logs when initializing the posting list, the first time it’s brought back into memory.

Queries

Let’s understand how query execution works, by looking at an example.

{
    me(func: uid(0x1)) {
      pred_A
      pred_B {
        pred_B1
        pred_B2
      }
      pred_C {
        pred_C1
        pred_C2 {
          pred_C21
      }
      }
  }
}

Let’s assume we have 3 Alpha instances, and instance id=2 receives this query. These are the steps:

• Send queries to look up keys = pred_A, 0x1, pred_B, 0x1, and pred_C, 0x1. These predicates could belong to 3 different groups, served by potentially different Alpha servers. So, this would typically incur at max 3 network calls (equal to number of predicates at this step).
• The above queries would return back 3 lists of UIDs or values. The result of pred_B and pred_C would be converted into queries for pred_Bi and pred_Ci.
• pred_Bi and pred_Ci would then cause at max 4 network calls, depending upon where these predicates are located. The keys for pred_Bi, for example, would be pred_Bi, res_pred_Bk, where res_pred_Bk = list of resulting UIDs from pred_B, u.
• Looking at res_pred_C2, you’ll notice that this would be a list of lists aka list matrix. We merge these list of lists into a sorted list with distinct elements to form the query for pred_C21.
• Another network call depending upon where pred_C21 lies, and this would again give us a list of list UIDs / value.

If the query was run via HTTP interface /query, this subgraph gets converted into JSON for replying back to the client. If the query was run via gRPC interface using the language clients, the subgraph gets converted to protocol buffer format and then returned to client.

Network Calls

Compared to RAM or SSD access, network calls are slow. Dgraph minimizes the number of network calls required to execute queries. As explained above, the data sharding is done based on predicate, not entity. Thus, even if we have a large set of intermediate results, they’d still only increase the payload of a network call, not the number of network calls itself. In general, the number of network calls done in Dgraph is directly proportional to the number of predicates in the query, or the complexity of the query, not the number of intermediate or final results.

In the above example, we have eight predicates, and so including a call to convert to UID, we’ll have at max nine network calls. The total number of entity results could be in millions.

Worker

In Queries section, you noticed how the calls were made to query for (predicate, uids). All those network calls / local processing are done via workers. Each server exposes a gRPC interface, which can then be called by the query processor to retrieve data.

Worker Pool

Worker Pool is just a pool of open TCP connections which can be reused by multiple goroutines. This avoids having to recreate a new connection every time a network call needs to be made.

Protocol Buffers

All data in Dgraph that is stored or transmitted is first converted into byte arrays through serialization using Protocol Buffers. When the result is to be returned to the user, the protocol buffer object is traversed, and the JSON object is formed.

Minimizing network calls explained

To explain how Dgraph minimizes network calls, let’s start with an example query we should be able to run.

Find all posts liked by friends of friends of mine over the last year, written by a popular author X.

SQL/NoSQL

In a distributed SQL/NoSQL database, this would require you to retrieve a lot of data.

Method 1:

• Find all the friends (~ 338 friends).
• Find all their friends (~ 338 * 338 = 40,000 people).
• Find all the posts liked by these people over the last year (resulting set in millions).
• Intersect these posts with posts authored by person X.

Method 2:

• Find all posts written by popular author X over the last year (possibly thousands).
• Find all people who liked those posts (easily millions) result set 1.
• Find all your friends.
• Find all their friends result set 2.
• Intersect result set 1 with result set 2.

Both of these approaches would result in a lot of data going back and forth between database and application; would be slow to execute, or would require you to run an offline job.

Dgraph

This is how it would run in Dgraph:

• Node X contains posting list for predicate friends.
• Seek to caller’s userid in Node X (1 RPC). Retrieve a list of friend uids.
• Do multiple seeks for each of the friend uids, to generate a list of friends of friends uids. result set 1
• Node Y contains posting list for predicate posts_liked.
• Ship result set 1 to Node Y (1 RPC), and do seeks to generate a list of all posts liked by result set 1. reult set 2
• Node Z contains posting list for predicate author.
• Ship result set 2 to Node Z (1 RPC). Seek to author X, and generate a list of posts authored by X. result set 3
• Intersect the two sorted lists, result set 2 and result set 3. result set 4
• Node N contains names for all uids.
• Ship result set 4 to Node N (1 RPC), and convert uids to names by doing multiple seeks. result set 5
• Ship result set 5 back to caller.

In 4-5 RPCs, we have figured out all the posts liked by friends of friends, written by popular author X.

This design allows vast scalability, and yet consistent production level latencies, to support running complicated queries requiring deep joins.

RAFT

This section aims to explain the RAFT consensus algorithm in simple terms. The idea is to give you just enough to make you understand the basic concepts, without going into explanations about why it works accurately. For a detailed explanation of RAFT, please read the original thesis paper by Diego Ongaro.

Term

Each election cycle is considered a term, during which there is a single leader (just like in a democracy). When a new election starts, the term number is increased. This is straightforward and obvious but is a critical factor for the accuracy of the algorithm.

In rare cases, if no leader could be elected within an ElectionTimeout, that term can end without a leader.

Server States

Each server in cluster can be in one of the following three states:

• Leader
• Follower
• Candidate

Generally, the servers are in leader or follower state. When the leader crashes or the communication breaks down, the followers will wait for election timeout before converting to candidates. The election timeout is randomized. This would allow one of them to declare candidacy before others. The candidate would vote for itself and wait for the majority of the cluster to vote for it as well. If a follower hears from a candidate with a higher term than the current (dead in this case) leader, it would vote for it. The candidate who gets majority votes wins the election and becomes the leader.

The leader then tells the rest of the cluster about the result (Heartbeat Communication) and the other candidates then become followers. Again, the cluster goes back into leader-follower model.

A leader could revert to being a follower without an election, if it finds another leader in the cluster with a higher Term). This might happen in rare cases (network partitions).

Communication

There is unidirectional RPC communication, from leader to followers. The followers never ping the leader. The leader sends AppendEntries messages to the followers with logs containing state updates. When the leader sends AppendEntries with zero logs, that’s considered a Heartbeat. Leader sends all followers Heartbeats at regular intervals.

If a follower doesn’t receive Heartbeat for ElectionTimeout duration (generally between 150ms to 300ms), it converts it’s state to candidate (as mentioned in Server States). It then requests for votes by sending a RequestVote call to other servers. Again, if it gets majority votes, candidate becomes a leader. At becoming leader, it then sends Heartbeats to all other servers to establish its authority (Cartman style, “Respect my authoritah!”).

Every communication request contains a term number. If a server receives a request with a stale term number, it rejects the request.

Raft believes in retrying RPCs indefinitely.

Log Entries

Log Entries are numbered sequentially and contain a term number. Entry is considered committed if it has been replicated to a majority of the servers.

On receiving a client request, the leader does four things (aka Log Replication):

• Appends and persists to its log.
• Issue AppendEntries in parallel to other servers.
• On majority replication, consider the entry committed and apply to its state machine.
• Notify followers that entry is committed so that they can apply it to their state machines.

A leader never overwrites or deletes its entries. There is a guarantee that if an entry is committed, all future leaders will have it. A leader can, however, force overwrite the followers’ logs, so they match leader’s logs (elected democratically, but got a dictator).

Voting

Each server persists its current term and vote, so it doesn’t end up voting twice in the same term. On receiving a RequestVote RPC, the server denies its vote if its log is more up-to-date than the candidate. It would also deny a vote, if a minimum ElectionTimeout hasn’t passed since the last Heartbeat from the leader. Otherwise, it gives a vote and resets its ElectionTimeout timer.

Up-to-date property of logs is determined as follows:

• Term number comparison
• Index number or log length comparison

Cluster membership

Raft only allows single-server changes, i.e. only one server can be added or deleted at a time. This is achieved by cluster configuration changes. Cluster configurations are communicated using special entries in AppendEntries.

The significant difference in how cluster configuration changes are applied compared to how typical Log Entries are applied is that the followers don’t wait for a commitment confirmation from the leader before enabling it.

A server can respond to both AppendEntries and RequestVote, without checking current configuration. This mechanism allows new servers to participate without officially being part of the cluster. Without this feature, things won’t work.

When a new server joins, it won’t have any logs, and they need to be streamed. To ensure cluster availability, Raft allows this server to join the cluster as a non-voting member. Once it’s caught up, voting can be enabled. This also allows the cluster to remove this server in case it’s too slow to catch up, before giving voting rights (sort of like getting a green card to allow assimilation before citizenship is awarded providing voting rights).

Log Compaction

One of the ways to do this is snapshotting. As soon as the state machine is synced to disk, the logs can be discarded.

Clients

Clients must locate the cluster to interact with it. Various approaches can be used for discovery.

A client can randomly pick up any server in the cluster. If the server isn’t a leader, the request should be rejected, and the leader information passed along. The client can then re-route it’s query to the leader. Alternatively, the server can proxy the client’s request to the leader.

When a client first starts up, it can register itself with the cluster using RegisterClient RPC. This creates a new client id, which is used for all subsequent RPCs.

Linearizable Semantics

Servers must filter out duplicate requests. They can do this via session tracking where they use the client id and another request UID set by the client to avoid reprocessing duplicate requests. RAFT also suggests storing responses along with the request UIDs to reply back in case it receives a duplicate request.

Linearizability requires the results of a read to reflect the latest committed write. Serializability, on the other hand, allows stale reads.

Read-only queries

To ensure linearizability of read-only queries run via leader, leader must take these steps:

• Leader must have at least one committed entry in its term. This would allow for up-to-dated-ness. (C’mon! Now that you’re in power do something at least!)
• Leader stores it’s latest commit index.
• Leader sends Heartbeats to the cluster and waits for ACK from majority. Now it knows that it’s the leader. (No successful coup. Yup, still the democratically elected dictator I was before!)
• Leader waits for its state machine to advance to readIndex.
• Leader can now run the queries against state machine and reply to clients.

Read-only queries can also be serviced by followers to reduce the load on the leader. But this could lead to stale results unless the follower confirms that its leader is the real leader(network partition). To do so, it would have to send a query to the leader, and the leader would have to do steps 1-3. Then the follower can do 4-5.

Read-only queries would have to be batched up, and then RPCs would have to go to the leader for each batch, who in turn would have to send further RPCs to the whole cluster. (This is not scalable without considerable optimizations to deal with latency.)

An alternative approach would be to have the servers return the index corresponding to their state machine. The client can then keep track of the maximum index it has received from replies so far. And pass it along to the server for the next request. If a server’s state machine hasn’t reached the index provided by the client, it will not service the request. This approach avoids inter-server communication and is a lot more scalable. (This approach does not guarantee linearizability, but should converge quickly to the latest write.)