tendermint/docs/tendermint-core/running-in-production.md

Running in production

Database

By default, Tendermint uses the syndtr/goleveldb package for it's in-process key-value database. Unfortunately, this implementation of LevelDB seems to suffer under heavy load (see #226). It may be best to install the real C-implementaiton of LevelDB and compile Tendermint to use that using make build_c. See the install instructions for details.

Tendermint keeps multiple distinct LevelDB databases in the $TMROOT/data:

• blockstore.db: Keeps the entire blockchain - stores blocks, block commits, and block meta data, each indexed by height. Used to sync new peers.
• evidence.db: Stores all verified evidence of misbehaviour.
• state.db: Stores the current blockchain state (ie. height, validators, consensus params). Only grows if consensus params or validators change. Also used to temporarily store intermediate results during block processing.
• tx_index.db: Indexes txs (and their results) by tx hash and by DeliverTx result tags.

By default, Tendermint will only index txs by their hash, not by their DeliverTx result tags. See indexing transactions for details.

There is no current strategy for pruning the databases. Consider reducing block production by controlling empty blocks or by increasing the consensus.timeout_commit param. Note both of these are local settings and not enforced by the consensus.

We're working on state syncing, which will enable history to be thrown away and recent application state to be directly synced. We'll need to develop solutions for archival nodes that allow queries on historical transactions and states. The Cosmos project has had much success just dumping the latest state of a blockchain to disk and starting a new chain from that state.

Logging

Default logging level (main:info,state:info,*:) should suffice for normal operation mode. Read this post for details on how to configure log_level config variable. Some of the modules can be found here. If you're trying to debug Tendermint or asked to provide logs with debug logging level, you can do so by running tendermint with --log_level="*:debug".

Write Ahead Logs (WAL)

Tendermint uses write ahead logs for the consensus (cs.wal) and the mempool (mempool.wal). Both WALs have a max size of 1GB and are automatically rotated.

Consensus WAL

The consensus.wal is used to ensure we can recover from a crash at any point in the consensus state machine. It writes all consensus messages (timeouts, proposals, block part, or vote) to a single file, flushing to disk before processing messages from its own validator. Since Tendermint validators are expected to never sign a conflicting vote, the WAL ensures we can always recover deterministically to the latest state of the consensus without using the network or re-signing any consensus messages.

If your consensus.wal is corrupted, see below.

Mempool WAL

The mempool.wal logs all incoming txs before running CheckTx, but is otherwise not used in any programmatic way. It's just a kind of manual safe guard. Note the mempool provides no durability guarantees - a tx sent to one or many nodes may never make it into the blockchain if those nodes crash before being able to propose it. Clients must monitor their txs by subscribing over websockets, polling for them, or using /broadcast_tx_commit. In the worst case, txs can be resent from the mempool WAL manually.

DOS Exposure and Mitigation

Validators are supposed to setup Sentry Node Architecture to prevent Denial-of-service attacks. You can read more about it here.

P2P

The core of the Tendermint peer-to-peer system is MConnection. Each connection has MaxPacketMsgPayloadSize, which is the maximum packet size and bounded send & receive queues. One can impose restrictions on send & receive rate per connection (SendRate, RecvRate).

RPC

Endpoints returning multiple entries are limited by default to return 30 elements (100 max). See the RPC Documentation for more information.

Rate-limiting and authentication are another key aspects to help protect against DOS attacks. While in the future we may implement these features, for now, validators are supposed to use external tools like NGINX or traefik to achieve the same things.

Debugging Tendermint

If you ever have to debug Tendermint, the first thing you should probably do is to check out the logs. See How to read logs, where we explain what certain log statements mean.

If, after skimming through the logs, things are not clear still, the next thing to try is query the /status RPC endpoint. It provides the necessary info: whenever the node is syncing or not, what height it is on, etc.

curl http(s)://{ip}:{rpcPort}/status

dump_consensus_state will give you a detailed overview of the consensus state (proposer, lastest validators, peers states). From it, you should be able to figure out why, for example, the network had halted.

curl http(s)://{ip}:{rpcPort}/dump_consensus_state

There is a reduced version of this endpoint - consensus_state, which returns just the votes seen at the current height.

• Github Issues
• StackOverflow questions

Monitoring Tendermint

Each Tendermint instance has a standard /health RPC endpoint, which responds with 200 (OK) if everything is fine and 500 (or no response) - if something is wrong.

Other useful endpoints include mentioned earlier /status, /net_info and /validators.

We have a small tool, called tm-monitor, which outputs information from the endpoints above plus some statistics. The tool can be found here.

Tendermint also can report and serve Prometheus metrics. See Metrics.

What happens when my app dies?

You are supposed to run Tendermint under a process supervisor (like systemd or runit). It will ensure Tendermint is always running (despite possible errors).

Getting back to the original question, if your application dies, Tendermint will panic. After a process supervisor restarts your application, Tendermint should be able to reconnect successfully. The order of restart does not matter for it.

Signal handling

We catch SIGINT and SIGTERM and try to clean up nicely. For other signals we use the default behaviour in Go: Default behavior of signals in Go programs.

Corruption

NOTE: Make sure you have a backup of the Tendermint data directory.

Possible causes

Remember that most corruption is caused by hardware issues:

• RAID controllers with faulty / worn out battery backup, and an unexpected power loss
• Hard disk drives with write-back cache enabled, and an unexpected power loss
• Cheap SSDs with insufficient power-loss protection, and an unexpected power-loss
• Defective RAM
• Defective or overheating CPU(s)

Other causes can be:

• Database systems configured with fsync=off and an OS crash or power loss
• Filesystems configured to use write barriers plus a storage layer that ignores write barriers. LVM is a particular culprit.
• Tendermint bugs
• Operating system bugs
• Admin error (e.g., directly modifying Tendermint data-directory contents)

(Source: https://wiki.postgresql.org/wiki/Corruption)

WAL Corruption

If consensus WAL is corrupted at the lastest height and you are trying to start Tendermint, replay will fail with panic.

Recovering from data corruption can be hard and time-consuming. Here are two approaches you can take:

1. Delete the WAL file and restart Tendermint. It will attempt to sync with other peers.
2. Try to repair the WAL file manually:

1. Create a backup of the corrupted WAL file:

cp "$TMHOME/data/cs.wal/wal" > /tmp/corrupted_wal_backup

2. Use ./scripts/wal2json to create a human-readable version

./scripts/wal2json/wal2json "$TMHOME/data/cs.wal/wal" > /tmp/corrupted_wal

3. Search for a "CORRUPTED MESSAGE" line.
4. By looking at the previous message and the message after the corrupted one and looking at the logs, try to rebuild the message. If the consequent messages are marked as corrupted too (this may happen if length header got corrupted or some writes did not make it to the WAL ~ truncation), then remove all the lines starting from the corrupted one and restart Tendermint.

$EDITOR /tmp/corrupted_wal

5. After editing, convert this file back into binary form by running:

./scripts/json2wal/json2wal /tmp/corrupted_wal  $TMHOME/data/cs.wal/wal

Hardware

Processor and Memory

While actual specs vary depending on the load and validators count, minimal requirements are:

• 1GB RAM
• 25GB of disk space
• 1.4 GHz CPU

SSD disks are preferable for applications with high transaction throughput.

Recommended:

• 2GB RAM
• 100GB SSD
• x64 2.0 GHz 2v CPU

While for now, Tendermint stores all the history and it may require significant disk space over time, we are planning to implement state syncing (See this issue). So, storing all the past blocks will not be necessary.

Operating Systems

Tendermint can be compiled for a wide range of operating systems thanks to Go language (the list of $OS/$ARCH pairs can be found here).

While we do not favor any operation system, more secure and stable Linux server distributions (like Centos) should be preferred over desktop operation systems (like Mac OS).

Miscellaneous

NOTE: if you are going to use Tendermint in a public domain, make sure you read hardware recommendations for a validator in the Cosmos network.

Configuration parameters

• p2p.flush_throttle_timeout
• p2p.max_packet_msg_payload_size
• p2p.send_rate
• p2p.recv_rate

If you are going to use Tendermint in a private domain and you have a private high-speed network among your peers, it makes sense to lower flush throttle timeout and increase other params.

[p2p]

send_rate=20000000 # 2MB/s
recv_rate=20000000 # 2MB/s
flush_throttle_timeout=10
max_packet_msg_payload_size=10240 # 10KB

• mempool.recheck

After every block, Tendermint rechecks every transaction left in the mempool to see if transactions committed in that block affected the application state, so some of the transactions left may become invalid. If that does not apply to your application, you can disable it by setting mempool.recheck=false.

• mempool.broadcast

Setting this to false will stop the mempool from relaying transactions to other peers until they are included in a block. It means only the peer you send the tx to will see it until it is included in a block.

• consensus.skip_timeout_commit

We want skip_timeout_commit=false when there is economics on the line because proposers should wait to hear for more votes. But if you don't care about that and want the fastest consensus, you can skip it. It will be kept false by default for public deployments (e.g. Cosmos Hub) while for enterprise applications, setting it to true is not a problem.

• consensus.peer_gossip_sleep_duration

You can try to reduce the time your node sleeps before checking if theres something to send its peers.

• consensus.timeout_commit

You can also try lowering timeout_commit (time we sleep before proposing the next block).

• p2p.addr_book_strict

By default, Tendermint checks whenever a peer's address is routable before saving it to the address book. The address is considered as routable if the IP is valid and within allowed ranges.

This may not be the case for private or local networks, where your IP range is usually strictly limited and private. If that case, you need to set addr_book_strict to false (turn it off).

• rpc.max_open_connections

By default, the number of simultaneous connections is limited because most OS give you limited number of file descriptors.

If you want to accept greater number of connections, you will need to increase these limits.

Sysctls to tune the system to be able to open more connections

...for N connections, such as 50k:

kern.maxfiles=10000+2*N         # BSD
kern.maxfilesperproc=100+2*N    # BSD
kern.ipc.maxsockets=10000+2*N   # BSD
fs.file-max=10000+2*N           # Linux
net.ipv4.tcp_max_orphans=N      # Linux

# For load-generating clients.
net.ipv4.ip_local_port_range="10000  65535"  # Linux.
net.inet.ip.portrange.first=10000  # BSD/Mac.
net.inet.ip.portrange.last=65535   # (Enough for N < 55535)
net.ipv4.tcp_tw_reuse=1         # Linux
net.inet.tcp.maxtcptw=2*N       # BSD

# If using netfilter on Linux:
net.netfilter.nf_conntrack_max=N
echo $((N/8)) > /sys/module/nf_conntrack/parameters/hashsize

The similar option exists for limiting the number of gRPC connections - rpc.grpc_max_open_connections.