Why crypto withdrawals get stuck: an operator's guide from approval to the chain
Why exchange withdrawals stall: approvals, hot-wallet float, nonces, RBF, TRON energy, Solana expiry, XRP tags, and how to retry without paying twice.
Kostiantyn Boiko, Blockchain Engineer
English original
A crypto withdrawal shows “pending” for one of two reasons. Either it is still inside your system, held by a limit, a screening check, an approval or an empty hot wallet, or it has left and the chain has not settled it. The first is a queue to work through. The second is money in flight, where the wrong retry pays the customer twice.
This is the outbound sequel to why USDT deposits get stuck. We run deposit, sweep and withdrawal pipelines on EVM chains, TRON, Solana, the XRP Ledger and Bitcoin in production, with custody approval policies in front of the signers. Every withdrawal passes nine stages:
- Request and limits
- Risk, AML and address screening
- Approval: policy, officers, whitelists, time locks
- Hot-wallet balance: float, and gas to move it
- Signing: nonce, fees, the transaction itself
- Broadcast
- Mempool
- Confirmation
- Ledger finalisation and notification
Stages 1 to 4 happen inside your perimeter. Stages 5 to 8 involve the chain, and each chain sets its own rules for a safe fix.
Where does a withdrawal wait before it reaches the chain?
Request and limits
The quiet failures are limits and holds shown as “pending” instead of a clear message:
- Daily caps per asset and customer tier. Refuse with the reason, or queue with a visible date.
- Security holds. Our custody policy holds new withdrawal addresses for 24 hours by default; our wallet holds withdrawals for 24 hours after a password reset. Both create tickets when the screen is silent.
- Limits computed from a price. On the reference stack, USD withdrawal limits come from a reference price, so a frozen or wrong price makes them wrong in either direction.
- Duplicate requests. A double tap or app retry creates two withdrawal rows; idempotency keys prevent it (below).
Give every held withdrawal a reason code and alert on counts per reason.
Risk, AML and address screening
- The screening vendor times out. Fail closed into a manual queue, and alert on vendor errors separately from queue age.
- The address is on a list. OFAC lists crypto addresses but says those listings are “not likely to be exhaustive”, so a list check is a minimum.
- The wrong network. A TRON address entered for an ERC-20 withdrawal fails a format check, but an EVM address is valid on every EVM chain. Make the customer pick the network.
Travel rule holds belong here too, and the law sets the behaviour:
- The requirement (EU). Article 14 of Regulation (EU) 2023/1113, applicable from 30 December 2024, requires the sender’s provider to send originator and beneficiary information with each crypto transfer and, for transfers above €1,000 to a self-hosted address, to assess whether the customer owns or controls that address. It “shall not allow for the initiation, or execute any transfer” until it complies.
- Our recommendation. Give “waiting for travel rule data” and “waiting for proof of address ownership” their own states and age alerts.
- Good practice. Collect the information on the withdrawal form, not by email afterwards.
Approval
Our custody defaults by amount band are 1 of 3 officers below a floor, 2 of 3 in the middle, and 3 of 5 plus a delay above a threshold, each approval a TOTP event by a named officer. Cold wallet moves need the full quorum and a time lock. We compared the schemes in MPC vs multisig for crypto exchanges.
Withdrawals stall here when the quorum cannot be reached, when nobody notices an approval waiting, or when an approved withdrawal is lost in the queue between approval and signer. Alert on “awaiting approval” against your service level and “approved but not signed” against minutes. Add approvers across time zones; do not lower thresholds under pressure.
Is there enough money in the hot wallet?
In our custody tiers the hot wallet pays withdrawals within a target balance and a ceiling per asset. When it runs dry, withdrawals queue behind a refill: warm-to-hot needs an operator signature plus 2FA, and cold needs the full quorum and a time lock. An hour of refill is an hour of “pending” for every customer.
Size the hot wallet float to cover the withdrawals you expect during your slowest realistic refill plus a margin, and alert below that level, not at zero. Route a withdrawal above the hot ceiling to a visible “large withdrawal” state.
The float can be there and the withdrawal still stuck:
- No gas. A USDT withdrawal needs ETH or TRX on the sending address. On our wallet stack the hot wallet and the fee wallet were the same address, which lets withdrawals drain the gas that sweeps need. Keep fee wallets separate.
- XRP reserves. An XRP Ledger account cannot send its reserve, currently 1 XRP plus 0.2 XRP per owned object on Mainnet. Paying into it fails with
tecUNFUNDED_PAYMENT; a payment too small to create a new account fails withtecNO_DST_INSUF_XRP. - Funds at a venue. In our exchange’s just-in-time cycle, withdrawals pull funds from the hedging venue, so venue transfer age needs its own alert.
What goes wrong at signing?
EVM nonces. The nonce is “a sequentially incrementing counter” per account; nonce n+1 cannot be mined before nonce n. In our production systems we saw parallel sweep transactions behind a load-balanced RPC pool read the same pending nonce, and one of each pair disappeared. The fix applies to withdrawals too: one nonce allocator per chain and sending address, shared by every signing worker, which reserves the nonce and stores the signed transaction before broadcast. An RPC timeout is not proof that the broadcast failed.
The opposite failure is a nonce gap: the counter advanced, the transaction never went out, and every later withdrawal waits behind it. Geth keeps such “queued” transactions for at most 3 hours by default (--txpool.lifetime), then drops them; that is a local node policy, not an on-chain expiry. Compare the confirmed nonce with your stored attempts, and fill a real gap with the missing nonce.
EVM fees. Under EIP-1559 a transaction sets maxFeePerGas and maxPriorityFeePerGas, and it waits while maxFeePerGas is below the block’s base fee. The base fee can rise by at most 12.5% per block, so a cap at twice the current base fee survives five full blocks in a row; the sixth takes it past the cap.
TRON energy. fee_limit caps the energy cost a transaction may burn. Beyond it the transaction fails with OUT_OF_ENERGY, and the energy used is not refunded. On 16 October 2026 a USDT transfer to an address holding zero USDT cost 13.03 TRX at 100 sun per energy unit, so a 10 TRX limit would fail those withdrawals and burn the 10 TRX; we use 15 (the numbers). Bandwidth is separate: an account gets 600 free bandwidth points a day, about one 345-byte USDT transfer, and beyond that TRX is burned, so a hot wallet on rented energy still needs TRX: energy rental does not cover bandwidth. Some TRON endpoints are API-only and omit fields a withdrawal needs, so our RPC gateway excludes them.
Solana fees. The base fee is 5,000 lamports per signature; the priority fee is the compute unit price in micro-lamports times the compute unit limit, divided by one million. Too low, and the transaction can expire unprocessed.
Token contracts that say no. The USDT contract on Ethereum checks whenNotPaused and require(!isBlackListed[msg.sender]) on every transfer: if the token is paused or your hot wallet is blacklisted, every USDT withdrawal reverts and still pays gas. Circle’s USDC contract also checks the recipient. Simulate each token withdrawal before signing (eth_call, or triggerconstantcontract on TRON) and alert on the revert reason instead of retrying. A simulation is a preflight check, not a guarantee: state can change before the transaction executes.
Why does a broadcast that was accepted disappear?
“Accepted” means one node took the transaction, nothing more.
EVM chains. An underpriced transaction stays pending until the base fee falls or a node evicts it. Replace it with the same nonce and higher fees: Geth’s default requires both the fee cap and the tip to rise by at least 10% (--txpool.pricebump), or it answers “replacement transaction underpriced”. “Nonce too low” means the nonce is used, possibly by your own earlier attempt.
Bitcoin. A low-feerate transaction waits, can be evicted when the mempool fills (300 MB by default in Bitcoin Core) and expires from a node’s mempool after 336 hours. Eviction is a node policy: it does not invalidate the transaction, and anyone holding it can rebroadcast it. BIP-125 defined replace-by-fee with opt-in signalling; Bitcoin Core 28.0 (October 2024) made full RBF the default, and 29.0 removed the option. A replacement must still pay at least the total fee it replaces plus the incremental relay fee for its own size: at a node setting of 0.1 sat/vB, 50 sats more for 500 vB. RBF changes the transaction ID, so update the hash the customer sees. CPFP, spending your change output in a high-fee child, leaves the original untouched.
TRON. A transaction built by a full node expires 60 seconds after the head block by default (24 hours at most) and is then dropped. It also references a recent block, and is rejected if that block is on a fork or outside the 65,536-block window. A node that is behind causes both.
Solana. A blockhash is valid for 150 slots, about 60 to 90 seconds, and RPC nodes rebroadcast every two seconds until the transaction is finalized or the blockhash expires. Fetch the blockhash at confirmed commitment (finalized costs about 13 seconds of the window), and beware lagging nodes, which can hand out a nearly expired one.
XRP Ledger. A preliminary tesSUCCESS is not final; only a validated ledger is. Set LastLedgerSequence on every transaction (the docs suggest four ledgers past the last validated one). Once the validated ledger index passes it, the transaction can no longer be included; check validated history over that range to learn whether it already succeeded.
RPC that disagrees with itself. One endpoint says pending, another has never heard of the transaction, a third is behind. Compare endpoint heights with the pool and confirm outcomes through a provider outside it before acting (RPC failover).
Confirmed on chain, but “USDT withdrawal not received”?
A hash is not a payment. A reverted EVM transaction is mined, uses its nonce and pays gas. A TRON transfer can be recorded with OUT_OF_ENERGY. A validated XRP transaction with a tec result charges the fee and consumes its sequence number. Mark a withdrawal paid from the result (token, recipient, amount), never from the hash. Our withdrawals have chain verification, and when we audited, every one marked successful was confirmed paid.
The destination needed a tag. An XRP payment to an exchange without a destination tag lands in its shared account for its support to find. Accounts that require a tag reject untagged payments with tecDST_TAG_NEEDED, but not every exchange sets that. Warn when the tag field is empty. The receiving side also has its own confirmations and minimums, so give the customer the hash and network to show them.
Depth and the ledger. Hold “completed” until your confirmation depth: Ethereum normally finalizes in about 15 minutes, but finality can stall, so check the finalized status rather than the clock; on TRON, wait for the solidified block. The dangerous gap is a withdrawal refunded while its transaction can still land: the customer withdraws again and is paid twice.
How do you retry a withdrawal without paying twice?
At the API, each withdrawal request carries a client-generated idempotency key. The IETF Idempotency-Key draft (version 08, October 2026) has the server answer a repeated key consistently with the original, return 409 while the first is still running, and 422 for the same key with a different payload. Stripe keeps keys for at least 24 hours; keep withdrawal keys for as long as you keep the withdrawal record.
Below the API, a retry must reuse whatever makes two attempts conflict on chain, so that at most one can confirm:
| Chain | What makes attempts conflict | Safe retry |
|---|---|---|
| Ethereum, BNB Chain, Avalanche | Same sender and nonce | Same nonce, both fees raised per the node’s replacement rule (Geth: at least 10%) |
| Bitcoin | Shared inputs | RBF spending the same inputs; CPFP on your change speeds up the original instead of replacing it |
| TRON | Identical signed bytes only | Rebroadcast the same signed bytes; rebuild only after expiry, with the ID absent from the chain |
| Solana | Identical transaction only | Rebroadcast the same transaction; re-sign only after lastValidBlockHeight |
| XRP Ledger | The account sequence | Same Sequence, new LastLedgerSequence and fee |
The Solana docs warn that if you re-sign while the first blockhash is valid, “it is possible for both transactions to be accepted by the network”.
In our treasury recovery runs, an unknown outcome locked the step until the chain said what happened. Apply that to every withdrawal: nobody resends and nobody refunds. “Failed” is proven only when the original can no longer land and did not already pay. It can no longer land when its expiry or LastLedgerSequence has passed, or when its nonce, sequence or inputs were used by another transaction; in that last case, find that transaction and check what it paid.
Stage by stage: symptom, cause, fix, alert
| Stage | Symptom | Likely cause | Fix | Alert to set |
|---|---|---|---|---|
| Request | Pending, no reason | Cap, hold, wrong price | Show reason and date | Held withdrawals by reason |
| Screening | Long review | Vendor timeout, travel rule data | Own states; data on the form | Vendor errors; queue age |
| Approval | Waiting for days | No quorum, forgotten | Approvers across time zones | Approval age; approved, not signed |
| Hot wallet | Everything pending | Float or gas empty, XRP reserve | Refill; separate fee wallet | Balance below refill-time demand |
| Signing | Nothing broadcast | Nonce gap, paused or blacklisted token | Fill the gap; simulate first | Counter vs chain nonce; reverts |
| Broadcast | Hash unknown to explorer | Dropped, expired, stale node | Safe retry (table above) | No receipt after expiry |
| Mempool | Pending for hours | Fee below market | Same-nonce replacement, RBF, CPFP | Fee below inclusion level |
| Confirmation | Completed, not received | Reverted, no tag, wrong network | Read the result; give hash and tag | Mined with failed status |
| Ledger | Paid but pending, or refunded but paid | Ledger and chain disagree | Reconcile; never refund an unknown | Ledger vs chain mismatch |
What should you tell the customer, and when?
- Show the real stage: “In review”, “Waiting for approval”, “Queued for funding”, “Sent, waiting for confirmations”, “Completed”. One “pending” for all of them creates most tickets.
- Before the chain, give a time window and say what you need. When a withdrawal passes your service level, tell the customer before they ask.
- Once broadcast, show the hash, network and an explorer link. After a replacement, show the new hash and explain why the old one shows as dropped.
- On chain but not received, give hash, network and tag for the receiving platform.
- Failed: confirm the refund only once the original provably cannot land.
Stuck withdrawals FAQ
Why is my crypto withdrawal pending on the exchange for hours?
Most long waits happen before the chain: a manual review, an approval quorum, a hold on a new address, or a hot wallet waiting for a refill. If the exchange has already given you a transaction hash, check its status on the correct network: a fee below the market is the most common on-chain cause.
How do I fix a stuck Ethereum transaction with a nonce gap?
Compare the confirmed nonce with your stored attempts, then send or replace the transaction at the first missing nonce; the transactions queued behind it become executable, unless they have fee or balance problems of their own. To speed up a pending transaction, resend the same nonce with both fee fields raised, by at least 10% on Geth’s defaults.
Can a stuck Bitcoin withdrawal be sped up?
Yes. The sender can replace it with a higher-fee version (RBF), which Bitcoin Core relays without opt-in signalling since version 28.0. The recipient, or the sender through its change output, can attach a high-fee child transaction (CPFP).
Why does my USDT withdrawal show completed but was not received?
Check the network first: USDT on TRON and USDT on Ethereum are separate tokens. If it is right, the receiving platform may still be confirming, may need a memo or tag, or may not credit amounts below its minimum, so send them the transaction hash.
Is it safe to cancel and resend a stuck withdrawal?
Only if the resend conflicts with the original: the same nonce on EVM chains, the same inputs on Bitcoin, the same sequence on the XRP Ledger, or after the original has expired on TRON and Solana. Before rebuilding or refunding after expiry, confirm that the original did not already pay; otherwise both can land and you pay twice.
How large should a hot wallet float be?
Large enough for the withdrawals you expect during your slowest realistic refill, including warm approvals and any cold time lock, plus a margin. Alert below that level, not at zero.
How BitPolaris helps
Custody runs these stages in our stack: nine chain bridges that build and broadcast withdrawals over RPC pools with failover; M-of-N approvals per asset and amount band, a 24-hour delay on new whitelist addresses and daily caps; hot, warm and cold tiers; separate fee wallets; checks that flag USDT TRC-20 versus ERC-20 mix-ups; and a stuck-money screen for transfers paid on chain that look wrong in the records, or waiting for a forgotten approval. Its reserve monitor reports a check it cannot complete as unknown, never as fine. Self-hosted, from €15,400 one-time.
The Exchange runs on the same custody core, with approvals, stuck funds and reconciliation in its operations console, and rents TRON energy at 4.7 to 9.3 TRX per withdrawal instead of burning about 13 TRX; if rental is down, withdrawals fall back to burning TRX instead of stopping. If withdrawals are your biggest ticket category, talk to us.
Sources
- Ethereum: EIP-1559, transactions and the nonce, time to finality
- Geth: command-line options, txpool namespace, source code for txpool errors, core errors and the replacement rule
- Bitcoin: BIP-125, Bitcoin Core 28.0 and 29.0 release notes, replacement policy, mempool defaults, Bitcoin Optech on CPFP
- TRON: setting fee_limit, resource model, transaction structure and expiration
- Solana: transactions, fee structure, confirmation and expiration, retrying transactions
- XRP Ledger: reserves, tec result codes, source and destination tags, reliable transaction submission
- Token contracts: Tether USDT on Ethereum, verified source, Circle FiatTokenV1 (USDC)
- Screening and travel rule: OFAC FAQ 562, Regulation (EU) 2023/1113
- Idempotency: IETF draft, The Idempotency-Key HTTP Header Field, Stripe idempotent requests